aoqi@0: /* aoqi@0: * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved. aoqi@0: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. aoqi@0: * aoqi@0: * This code is free software; you can redistribute it and/or modify it aoqi@0: * under the terms of the GNU General Public License version 2 only, as aoqi@0: * published by the Free Software Foundation. aoqi@0: * aoqi@0: * This code is distributed in the hope that it will be useful, but WITHOUT aoqi@0: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or aoqi@0: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License aoqi@0: * version 2 for more details (a copy is included in the LICENSE file that aoqi@0: * accompanied this code). aoqi@0: * aoqi@0: * You should have received a copy of the GNU General Public License version aoqi@0: * 2 along with this work; if not, write to the Free Software Foundation, aoqi@0: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. aoqi@0: * aoqi@0: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA aoqi@0: * or visit www.oracle.com if you need additional information or have any aoqi@0: * questions. aoqi@0: * aoqi@0: */ aoqi@0: aoqi@0: #include "precompiled.hpp" aoqi@0: #include "classfile/systemDictionary.hpp" aoqi@0: #include "classfile/vmSymbols.hpp" aoqi@0: #include "compiler/compileBroker.hpp" aoqi@0: #include "compiler/compileLog.hpp" aoqi@0: #include "oops/objArrayKlass.hpp" aoqi@0: #include "opto/addnode.hpp" aoqi@0: #include "opto/callGenerator.hpp" aoqi@0: #include "opto/cfgnode.hpp" aoqi@0: #include "opto/idealKit.hpp" aoqi@0: #include "opto/mathexactnode.hpp" aoqi@0: #include "opto/mulnode.hpp" aoqi@0: #include "opto/parse.hpp" aoqi@0: #include "opto/runtime.hpp" aoqi@0: #include "opto/subnode.hpp" aoqi@0: #include "prims/nativeLookup.hpp" aoqi@0: #include "runtime/sharedRuntime.hpp" aoqi@0: #include "trace/traceMacros.hpp" aoqi@0: aoqi@0: class LibraryIntrinsic : public InlineCallGenerator { aoqi@0: // Extend the set of intrinsics known to the runtime: aoqi@0: public: aoqi@0: private: aoqi@0: bool _is_virtual; aoqi@0: bool _is_predicted; aoqi@0: bool _does_virtual_dispatch; aoqi@0: vmIntrinsics::ID _intrinsic_id; aoqi@0: aoqi@0: public: aoqi@0: LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id) aoqi@0: : InlineCallGenerator(m), aoqi@0: _is_virtual(is_virtual), aoqi@0: _is_predicted(is_predicted), aoqi@0: _does_virtual_dispatch(does_virtual_dispatch), aoqi@0: _intrinsic_id(id) aoqi@0: { aoqi@0: } aoqi@0: virtual bool is_intrinsic() const { return true; } aoqi@0: virtual bool is_virtual() const { return _is_virtual; } aoqi@0: virtual bool is_predicted() const { return _is_predicted; } aoqi@0: virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } aoqi@0: virtual JVMState* generate(JVMState* jvms, Parse* parent_parser); aoqi@0: virtual Node* generate_predicate(JVMState* jvms); aoqi@0: vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } aoqi@0: }; aoqi@0: aoqi@0: aoqi@0: // Local helper class for LibraryIntrinsic: aoqi@0: class LibraryCallKit : public GraphKit { aoqi@0: private: aoqi@0: LibraryIntrinsic* _intrinsic; // the library intrinsic being called aoqi@0: Node* _result; // the result node, if any aoqi@0: int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted aoqi@0: aoqi@0: const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); aoqi@0: aoqi@0: public: aoqi@0: LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) aoqi@0: : GraphKit(jvms), aoqi@0: _intrinsic(intrinsic), aoqi@0: _result(NULL) aoqi@0: { aoqi@0: // Check if this is a root compile. In that case we don't have a caller. aoqi@0: if (!jvms->has_method()) { aoqi@0: _reexecute_sp = sp(); aoqi@0: } else { aoqi@0: // Find out how many arguments the interpreter needs when deoptimizing aoqi@0: // and save the stack pointer value so it can used by uncommon_trap. aoqi@0: // We find the argument count by looking at the declared signature. aoqi@0: bool ignored_will_link; aoqi@0: ciSignature* declared_signature = NULL; aoqi@0: ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); aoqi@0: const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); aoqi@0: _reexecute_sp = sp() + nargs; // "push" arguments back on stack aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } aoqi@0: aoqi@0: ciMethod* caller() const { return jvms()->method(); } aoqi@0: int bci() const { return jvms()->bci(); } aoqi@0: LibraryIntrinsic* intrinsic() const { return _intrinsic; } aoqi@0: vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } aoqi@0: ciMethod* callee() const { return _intrinsic->method(); } aoqi@0: aoqi@0: bool try_to_inline(); aoqi@0: Node* try_to_predicate(); aoqi@0: aoqi@0: void push_result() { aoqi@0: // Push the result onto the stack. aoqi@0: if (!stopped() && result() != NULL) { aoqi@0: BasicType bt = result()->bottom_type()->basic_type(); aoqi@0: push_node(bt, result()); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: private: aoqi@0: void fatal_unexpected_iid(vmIntrinsics::ID iid) { aoqi@0: fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); aoqi@0: } aoqi@0: aoqi@0: void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } aoqi@0: void set_result(RegionNode* region, PhiNode* value); aoqi@0: Node* result() { return _result; } aoqi@0: aoqi@0: virtual int reexecute_sp() { return _reexecute_sp; } aoqi@0: aoqi@0: // Helper functions to inline natives aoqi@0: Node* generate_guard(Node* test, RegionNode* region, float true_prob); aoqi@0: Node* generate_slow_guard(Node* test, RegionNode* region); aoqi@0: Node* generate_fair_guard(Node* test, RegionNode* region); aoqi@0: Node* generate_negative_guard(Node* index, RegionNode* region, aoqi@0: // resulting CastII of index: aoqi@0: Node* *pos_index = NULL); aoqi@0: Node* generate_nonpositive_guard(Node* index, bool never_negative, aoqi@0: // resulting CastII of index: aoqi@0: Node* *pos_index = NULL); aoqi@0: Node* generate_limit_guard(Node* offset, Node* subseq_length, aoqi@0: Node* array_length, aoqi@0: RegionNode* region); aoqi@0: Node* generate_current_thread(Node* &tls_output); aoqi@0: address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, aoqi@0: bool disjoint_bases, const char* &name, bool dest_uninitialized); aoqi@0: Node* load_mirror_from_klass(Node* klass); aoqi@0: Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, aoqi@0: RegionNode* region, int null_path, aoqi@0: int offset); aoqi@0: Node* load_klass_from_mirror(Node* mirror, bool never_see_null, aoqi@0: RegionNode* region, int null_path) { aoqi@0: int offset = java_lang_Class::klass_offset_in_bytes(); aoqi@0: return load_klass_from_mirror_common(mirror, never_see_null, aoqi@0: region, null_path, aoqi@0: offset); aoqi@0: } aoqi@0: Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, aoqi@0: RegionNode* region, int null_path) { aoqi@0: int offset = java_lang_Class::array_klass_offset_in_bytes(); aoqi@0: return load_klass_from_mirror_common(mirror, never_see_null, aoqi@0: region, null_path, aoqi@0: offset); aoqi@0: } aoqi@0: Node* generate_access_flags_guard(Node* kls, aoqi@0: int modifier_mask, int modifier_bits, aoqi@0: RegionNode* region); aoqi@0: Node* generate_interface_guard(Node* kls, RegionNode* region); aoqi@0: Node* generate_array_guard(Node* kls, RegionNode* region) { aoqi@0: return generate_array_guard_common(kls, region, false, false); aoqi@0: } aoqi@0: Node* generate_non_array_guard(Node* kls, RegionNode* region) { aoqi@0: return generate_array_guard_common(kls, region, false, true); aoqi@0: } aoqi@0: Node* generate_objArray_guard(Node* kls, RegionNode* region) { aoqi@0: return generate_array_guard_common(kls, region, true, false); aoqi@0: } aoqi@0: Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { aoqi@0: return generate_array_guard_common(kls, region, true, true); aoqi@0: } aoqi@0: Node* generate_array_guard_common(Node* kls, RegionNode* region, aoqi@0: bool obj_array, bool not_array); aoqi@0: Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); aoqi@0: CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, aoqi@0: bool is_virtual = false, bool is_static = false); aoqi@0: CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { aoqi@0: return generate_method_call(method_id, false, true); aoqi@0: } aoqi@0: CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { aoqi@0: return generate_method_call(method_id, true, false); aoqi@0: } aoqi@0: Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); aoqi@0: aoqi@0: Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); aoqi@0: Node* make_string_method_node(int opcode, Node* str1, Node* str2); aoqi@0: bool inline_string_compareTo(); aoqi@0: bool inline_string_indexOf(); aoqi@0: Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); aoqi@0: bool inline_string_equals(); aoqi@0: Node* round_double_node(Node* n); aoqi@0: bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); aoqi@0: bool inline_math_native(vmIntrinsics::ID id); aoqi@0: bool inline_trig(vmIntrinsics::ID id); aoqi@0: bool inline_math(vmIntrinsics::ID id); aoqi@0: template aoqi@0: bool inline_math_overflow(Node* arg1, Node* arg2); aoqi@0: void inline_math_mathExact(Node* math, Node* test); aoqi@0: bool inline_math_addExactI(bool is_increment); aoqi@0: bool inline_math_addExactL(bool is_increment); aoqi@0: bool inline_math_multiplyExactI(); aoqi@0: bool inline_math_multiplyExactL(); aoqi@0: bool inline_math_negateExactI(); aoqi@0: bool inline_math_negateExactL(); aoqi@0: bool inline_math_subtractExactI(bool is_decrement); aoqi@0: bool inline_math_subtractExactL(bool is_decrement); aoqi@0: bool inline_exp(); aoqi@0: bool inline_pow(); aoqi@0: Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); aoqi@0: bool inline_min_max(vmIntrinsics::ID id); aoqi@0: Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); aoqi@0: // This returns Type::AnyPtr, RawPtr, or OopPtr. aoqi@0: int classify_unsafe_addr(Node* &base, Node* &offset); aoqi@0: Node* make_unsafe_address(Node* base, Node* offset); aoqi@0: // Helper for inline_unsafe_access. aoqi@0: // Generates the guards that check whether the result of aoqi@0: // Unsafe.getObject should be recorded in an SATB log buffer. aoqi@0: void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); aoqi@0: bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); aoqi@0: bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); aoqi@0: static bool klass_needs_init_guard(Node* kls); aoqi@0: bool inline_unsafe_allocate(); aoqi@0: bool inline_unsafe_copyMemory(); aoqi@0: bool inline_native_currentThread(); aoqi@0: #ifdef TRACE_HAVE_INTRINSICS aoqi@0: bool inline_native_classID(); aoqi@0: bool inline_native_threadID(); aoqi@0: #endif aoqi@0: bool inline_native_time_funcs(address method, const char* funcName); aoqi@0: bool inline_native_isInterrupted(); aoqi@0: bool inline_native_Class_query(vmIntrinsics::ID id); aoqi@0: bool inline_native_subtype_check(); aoqi@0: aoqi@0: bool inline_native_newArray(); aoqi@0: bool inline_native_getLength(); aoqi@0: bool inline_array_copyOf(bool is_copyOfRange); aoqi@0: bool inline_array_equals(); aoqi@0: void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); aoqi@0: bool inline_native_clone(bool is_virtual); aoqi@0: bool inline_native_Reflection_getCallerClass(); aoqi@0: // Helper function for inlining native object hash method aoqi@0: bool inline_native_hashcode(bool is_virtual, bool is_static); aoqi@0: bool inline_native_getClass(); aoqi@0: aoqi@0: // Helper functions for inlining arraycopy aoqi@0: bool inline_arraycopy(); aoqi@0: void generate_arraycopy(const TypePtr* adr_type, aoqi@0: BasicType basic_elem_type, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* copy_length, aoqi@0: bool disjoint_bases = false, aoqi@0: bool length_never_negative = false, aoqi@0: RegionNode* slow_region = NULL); aoqi@0: AllocateArrayNode* tightly_coupled_allocation(Node* ptr, aoqi@0: RegionNode* slow_region); aoqi@0: void generate_clear_array(const TypePtr* adr_type, aoqi@0: Node* dest, aoqi@0: BasicType basic_elem_type, aoqi@0: Node* slice_off, aoqi@0: Node* slice_len, aoqi@0: Node* slice_end); aoqi@0: bool generate_block_arraycopy(const TypePtr* adr_type, aoqi@0: BasicType basic_elem_type, aoqi@0: AllocateNode* alloc, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* dest_size, bool dest_uninitialized); aoqi@0: void generate_slow_arraycopy(const TypePtr* adr_type, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* copy_length, bool dest_uninitialized); aoqi@0: Node* generate_checkcast_arraycopy(const TypePtr* adr_type, aoqi@0: Node* dest_elem_klass, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* copy_length, bool dest_uninitialized); aoqi@0: Node* generate_generic_arraycopy(const TypePtr* adr_type, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* copy_length, bool dest_uninitialized); aoqi@0: void generate_unchecked_arraycopy(const TypePtr* adr_type, aoqi@0: BasicType basic_elem_type, aoqi@0: bool disjoint_bases, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* copy_length, bool dest_uninitialized); aoqi@0: typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; aoqi@0: bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); aoqi@0: bool inline_unsafe_ordered_store(BasicType type); aoqi@0: bool inline_unsafe_fence(vmIntrinsics::ID id); aoqi@0: bool inline_fp_conversions(vmIntrinsics::ID id); aoqi@0: bool inline_number_methods(vmIntrinsics::ID id); aoqi@0: bool inline_reference_get(); aoqi@0: bool inline_aescrypt_Block(vmIntrinsics::ID id); aoqi@0: bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); aoqi@0: Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); aoqi@0: Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); aoqi@0: Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); aoqi@0: bool inline_encodeISOArray(); aoqi@0: bool inline_updateCRC32(); aoqi@0: bool inline_updateBytesCRC32(); aoqi@0: bool inline_updateByteBufferCRC32(); aoqi@0: }; aoqi@0: aoqi@0: aoqi@0: //---------------------------make_vm_intrinsic---------------------------- aoqi@0: CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { aoqi@0: vmIntrinsics::ID id = m->intrinsic_id(); aoqi@0: assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); aoqi@0: aoqi@0: if (DisableIntrinsic[0] != '\0' aoqi@0: && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { aoqi@0: // disabled by a user request on the command line: aoqi@0: // example: -XX:DisableIntrinsic=_hashCode,_getClass aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: if (!m->is_loaded()) { aoqi@0: // do not attempt to inline unloaded methods aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: // Only a few intrinsics implement a virtual dispatch. aoqi@0: // They are expensive calls which are also frequently overridden. aoqi@0: if (is_virtual) { aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_hashCode: aoqi@0: case vmIntrinsics::_clone: aoqi@0: // OK, Object.hashCode and Object.clone intrinsics come in both flavors aoqi@0: break; aoqi@0: default: aoqi@0: return NULL; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // -XX:-InlineNatives disables nearly all intrinsics: aoqi@0: if (!InlineNatives) { aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_indexOf: aoqi@0: case vmIntrinsics::_compareTo: aoqi@0: case vmIntrinsics::_equals: aoqi@0: case vmIntrinsics::_equalsC: aoqi@0: case vmIntrinsics::_getAndAddInt: aoqi@0: case vmIntrinsics::_getAndAddLong: aoqi@0: case vmIntrinsics::_getAndSetInt: aoqi@0: case vmIntrinsics::_getAndSetLong: aoqi@0: case vmIntrinsics::_getAndSetObject: aoqi@0: case vmIntrinsics::_loadFence: aoqi@0: case vmIntrinsics::_storeFence: aoqi@0: case vmIntrinsics::_fullFence: aoqi@0: break; // InlineNatives does not control String.compareTo aoqi@0: case vmIntrinsics::_Reference_get: aoqi@0: break; // InlineNatives does not control Reference.get aoqi@0: default: aoqi@0: return NULL; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: bool is_predicted = false; aoqi@0: bool does_virtual_dispatch = false; aoqi@0: aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_compareTo: aoqi@0: if (!SpecialStringCompareTo) return NULL; aoqi@0: if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_indexOf: aoqi@0: if (!SpecialStringIndexOf) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_equals: aoqi@0: if (!SpecialStringEquals) return NULL; aoqi@0: if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_equalsC: aoqi@0: if (!SpecialArraysEquals) return NULL; aoqi@0: if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_arraycopy: aoqi@0: if (!InlineArrayCopy) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_copyMemory: aoqi@0: if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; aoqi@0: if (!InlineArrayCopy) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_hashCode: aoqi@0: if (!InlineObjectHash) return NULL; aoqi@0: does_virtual_dispatch = true; aoqi@0: break; aoqi@0: case vmIntrinsics::_clone: aoqi@0: does_virtual_dispatch = true; aoqi@0: case vmIntrinsics::_copyOf: aoqi@0: case vmIntrinsics::_copyOfRange: aoqi@0: if (!InlineObjectCopy) return NULL; aoqi@0: // These also use the arraycopy intrinsic mechanism: aoqi@0: if (!InlineArrayCopy) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_encodeISOArray: aoqi@0: if (!SpecialEncodeISOArray) return NULL; aoqi@0: if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_checkIndex: aoqi@0: // We do not intrinsify this. The optimizer does fine with it. aoqi@0: return NULL; aoqi@0: aoqi@0: case vmIntrinsics::_getCallerClass: aoqi@0: if (!UseNewReflection) return NULL; aoqi@0: if (!InlineReflectionGetCallerClass) return NULL; aoqi@0: if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_bitCount_i: aoqi@0: if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_bitCount_l: aoqi@0: if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_numberOfLeadingZeros_i: aoqi@0: if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_numberOfLeadingZeros_l: aoqi@0: if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_numberOfTrailingZeros_i: aoqi@0: if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_numberOfTrailingZeros_l: aoqi@0: if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_reverseBytes_c: aoqi@0: if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_reverseBytes_s: aoqi@0: if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_reverseBytes_i: aoqi@0: if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_reverseBytes_l: aoqi@0: if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_Reference_get: aoqi@0: // Use the intrinsic version of Reference.get() so that the value in aoqi@0: // the referent field can be registered by the G1 pre-barrier code. aoqi@0: // Also add memory barrier to prevent commoning reads from this field aoqi@0: // across safepoint since GC can change it value. aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_compareAndSwapObject: aoqi@0: #ifdef _LP64 aoqi@0: if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; aoqi@0: #endif aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_compareAndSwapLong: aoqi@0: if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_getAndAddInt: aoqi@0: if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_getAndAddLong: aoqi@0: if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_getAndSetInt: aoqi@0: if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_getAndSetLong: aoqi@0: if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_getAndSetObject: aoqi@0: #ifdef _LP64 aoqi@0: if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; aoqi@0: if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; aoqi@0: break; aoqi@0: #else aoqi@0: if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; aoqi@0: break; aoqi@0: #endif aoqi@0: aoqi@0: case vmIntrinsics::_aescrypt_encryptBlock: aoqi@0: case vmIntrinsics::_aescrypt_decryptBlock: aoqi@0: if (!UseAESIntrinsics) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: aoqi@0: case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: aoqi@0: if (!UseAESIntrinsics) return NULL; aoqi@0: // these two require the predicated logic aoqi@0: is_predicted = true; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_updateCRC32: aoqi@0: case vmIntrinsics::_updateBytesCRC32: aoqi@0: case vmIntrinsics::_updateByteBufferCRC32: aoqi@0: if (!UseCRC32Intrinsics) return NULL; aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_incrementExactI: aoqi@0: case vmIntrinsics::_addExactI: aoqi@0: if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_incrementExactL: aoqi@0: case vmIntrinsics::_addExactL: aoqi@0: if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_decrementExactI: aoqi@0: case vmIntrinsics::_subtractExactI: aoqi@0: if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_decrementExactL: aoqi@0: case vmIntrinsics::_subtractExactL: aoqi@0: if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_negateExactI: aoqi@0: if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_negateExactL: aoqi@0: if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_multiplyExactI: aoqi@0: if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; aoqi@0: break; aoqi@0: case vmIntrinsics::_multiplyExactL: aoqi@0: if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; aoqi@0: break; aoqi@0: aoqi@0: default: aoqi@0: assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); aoqi@0: assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); aoqi@0: break; aoqi@0: } aoqi@0: aoqi@0: // -XX:-InlineClassNatives disables natives from the Class class. aoqi@0: // The flag applies to all reflective calls, notably Array.newArray aoqi@0: // (visible to Java programmers as Array.newInstance). aoqi@0: if (m->holder()->name() == ciSymbol::java_lang_Class() || aoqi@0: m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { aoqi@0: if (!InlineClassNatives) return NULL; aoqi@0: } aoqi@0: aoqi@0: // -XX:-InlineThreadNatives disables natives from the Thread class. aoqi@0: if (m->holder()->name() == ciSymbol::java_lang_Thread()) { aoqi@0: if (!InlineThreadNatives) return NULL; aoqi@0: } aoqi@0: aoqi@0: // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. aoqi@0: if (m->holder()->name() == ciSymbol::java_lang_Math() || aoqi@0: m->holder()->name() == ciSymbol::java_lang_Float() || aoqi@0: m->holder()->name() == ciSymbol::java_lang_Double()) { aoqi@0: if (!InlineMathNatives) return NULL; aoqi@0: } aoqi@0: aoqi@0: // -XX:-InlineUnsafeOps disables natives from the Unsafe class. aoqi@0: if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { aoqi@0: if (!InlineUnsafeOps) return NULL; aoqi@0: } aoqi@0: aoqi@0: return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id); aoqi@0: } aoqi@0: aoqi@0: //----------------------register_library_intrinsics----------------------- aoqi@0: // Initialize this file's data structures, for each Compile instance. aoqi@0: void Compile::register_library_intrinsics() { aoqi@0: // Nothing to do here. aoqi@0: } aoqi@0: aoqi@0: JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) { aoqi@0: LibraryCallKit kit(jvms, this); aoqi@0: Compile* C = kit.C; aoqi@0: int nodes = C->unique(); aoqi@0: #ifndef PRODUCT aoqi@0: if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { aoqi@0: char buf[1000]; aoqi@0: const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); aoqi@0: tty->print_cr("Intrinsic %s", str); aoqi@0: } aoqi@0: #endif aoqi@0: ciMethod* callee = kit.callee(); aoqi@0: const int bci = kit.bci(); aoqi@0: aoqi@0: // Try to inline the intrinsic. aoqi@0: if (kit.try_to_inline()) { aoqi@0: if (C->print_intrinsics() || C->print_inlining()) { aoqi@0: C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); aoqi@0: } aoqi@0: C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); aoqi@0: if (C->log()) { aoqi@0: C->log()->elem("intrinsic id='%s'%s nodes='%d'", aoqi@0: vmIntrinsics::name_at(intrinsic_id()), aoqi@0: (is_virtual() ? " virtual='1'" : ""), aoqi@0: C->unique() - nodes); aoqi@0: } aoqi@0: // Push the result from the inlined method onto the stack. aoqi@0: kit.push_result(); aoqi@0: return kit.transfer_exceptions_into_jvms(); aoqi@0: } aoqi@0: aoqi@0: // The intrinsic bailed out aoqi@0: if (C->print_intrinsics() || C->print_inlining()) { aoqi@0: if (jvms->has_method()) { aoqi@0: // Not a root compile. aoqi@0: const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; aoqi@0: C->print_inlining(callee, jvms->depth() - 1, bci, msg); aoqi@0: } else { aoqi@0: // Root compile aoqi@0: tty->print("Did not generate intrinsic %s%s at bci:%d in", aoqi@0: vmIntrinsics::name_at(intrinsic_id()), aoqi@0: (is_virtual() ? " (virtual)" : ""), bci); aoqi@0: } aoqi@0: } aoqi@0: C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) { aoqi@0: LibraryCallKit kit(jvms, this); aoqi@0: Compile* C = kit.C; aoqi@0: int nodes = C->unique(); aoqi@0: #ifndef PRODUCT aoqi@0: assert(is_predicted(), "sanity"); aoqi@0: if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { aoqi@0: char buf[1000]; aoqi@0: const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); aoqi@0: tty->print_cr("Predicate for intrinsic %s", str); aoqi@0: } aoqi@0: #endif aoqi@0: ciMethod* callee = kit.callee(); aoqi@0: const int bci = kit.bci(); aoqi@0: aoqi@0: Node* slow_ctl = kit.try_to_predicate(); aoqi@0: if (!kit.failing()) { aoqi@0: if (C->print_intrinsics() || C->print_inlining()) { aoqi@0: C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); aoqi@0: } aoqi@0: C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); aoqi@0: if (C->log()) { aoqi@0: C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", aoqi@0: vmIntrinsics::name_at(intrinsic_id()), aoqi@0: (is_virtual() ? " virtual='1'" : ""), aoqi@0: C->unique() - nodes); aoqi@0: } aoqi@0: return slow_ctl; // Could be NULL if the check folds. aoqi@0: } aoqi@0: aoqi@0: // The intrinsic bailed out aoqi@0: if (C->print_intrinsics() || C->print_inlining()) { aoqi@0: if (jvms->has_method()) { aoqi@0: // Not a root compile. aoqi@0: const char* msg = "failed to generate predicate for intrinsic"; aoqi@0: C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); aoqi@0: } else { aoqi@0: // Root compile aoqi@0: C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", aoqi@0: vmIntrinsics::name_at(intrinsic_id()), aoqi@0: (is_virtual() ? " (virtual)" : ""), bci); aoqi@0: } aoqi@0: } aoqi@0: C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::try_to_inline() { aoqi@0: // Handle symbolic names for otherwise undistinguished boolean switches: aoqi@0: const bool is_store = true; aoqi@0: const bool is_native_ptr = true; aoqi@0: const bool is_static = true; aoqi@0: const bool is_volatile = true; aoqi@0: aoqi@0: if (!jvms()->has_method()) { aoqi@0: // Root JVMState has a null method. aoqi@0: assert(map()->memory()->Opcode() == Op_Parm, ""); aoqi@0: // Insert the memory aliasing node aoqi@0: set_all_memory(reset_memory()); aoqi@0: } aoqi@0: assert(merged_memory(), ""); aoqi@0: aoqi@0: aoqi@0: switch (intrinsic_id()) { aoqi@0: case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); aoqi@0: case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); aoqi@0: case vmIntrinsics::_getClass: return inline_native_getClass(); aoqi@0: aoqi@0: case vmIntrinsics::_dsin: aoqi@0: case vmIntrinsics::_dcos: aoqi@0: case vmIntrinsics::_dtan: aoqi@0: case vmIntrinsics::_dabs: aoqi@0: case vmIntrinsics::_datan2: aoqi@0: case vmIntrinsics::_dsqrt: aoqi@0: case vmIntrinsics::_dexp: aoqi@0: case vmIntrinsics::_dlog: aoqi@0: case vmIntrinsics::_dlog10: aoqi@0: case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); aoqi@0: aoqi@0: case vmIntrinsics::_min: aoqi@0: case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); aoqi@0: aoqi@0: case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); aoqi@0: case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); aoqi@0: case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); aoqi@0: case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); aoqi@0: case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); aoqi@0: case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); aoqi@0: case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); aoqi@0: case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); aoqi@0: case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); aoqi@0: case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); aoqi@0: case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); aoqi@0: case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); aoqi@0: aoqi@0: case vmIntrinsics::_arraycopy: return inline_arraycopy(); aoqi@0: aoqi@0: case vmIntrinsics::_compareTo: return inline_string_compareTo(); aoqi@0: case vmIntrinsics::_indexOf: return inline_string_indexOf(); aoqi@0: case vmIntrinsics::_equals: return inline_string_equals(); aoqi@0: aoqi@0: case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); aoqi@0: case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); aoqi@0: case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); aoqi@0: case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); aoqi@0: case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); aoqi@0: case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); aoqi@0: case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); aoqi@0: case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); aoqi@0: case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); aoqi@0: aoqi@0: case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); aoqi@0: case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); aoqi@0: case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); aoqi@0: case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); aoqi@0: case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); aoqi@0: case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); aoqi@0: case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); aoqi@0: case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); aoqi@0: case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); aoqi@0: aoqi@0: case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); aoqi@0: case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); aoqi@0: case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); aoqi@0: case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); aoqi@0: case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); aoqi@0: case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); aoqi@0: case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); aoqi@0: case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); aoqi@0: aoqi@0: case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); aoqi@0: case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); aoqi@0: case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); aoqi@0: case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); aoqi@0: case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); aoqi@0: case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); aoqi@0: case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); aoqi@0: case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); aoqi@0: aoqi@0: case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); aoqi@0: case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); aoqi@0: case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); aoqi@0: case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); aoqi@0: case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); aoqi@0: case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); aoqi@0: case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); aoqi@0: case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); aoqi@0: case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); aoqi@0: aoqi@0: case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); aoqi@0: case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); aoqi@0: case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); aoqi@0: case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); aoqi@0: case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); aoqi@0: case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); aoqi@0: case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); aoqi@0: case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); aoqi@0: case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); aoqi@0: aoqi@0: case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); aoqi@0: case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); aoqi@0: case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); aoqi@0: case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); aoqi@0: aoqi@0: case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); aoqi@0: case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); aoqi@0: case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); aoqi@0: aoqi@0: case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); aoqi@0: case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); aoqi@0: case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); aoqi@0: aoqi@0: case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); aoqi@0: case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); aoqi@0: case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); aoqi@0: case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); aoqi@0: case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); aoqi@0: aoqi@0: case vmIntrinsics::_loadFence: aoqi@0: case vmIntrinsics::_storeFence: aoqi@0: case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); aoqi@0: aoqi@0: case vmIntrinsics::_currentThread: return inline_native_currentThread(); aoqi@0: case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); aoqi@0: aoqi@0: #ifdef TRACE_HAVE_INTRINSICS aoqi@0: case vmIntrinsics::_classID: return inline_native_classID(); aoqi@0: case vmIntrinsics::_threadID: return inline_native_threadID(); aoqi@0: case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); aoqi@0: #endif aoqi@0: case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); aoqi@0: case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); aoqi@0: case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); aoqi@0: case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); aoqi@0: case vmIntrinsics::_newArray: return inline_native_newArray(); aoqi@0: case vmIntrinsics::_getLength: return inline_native_getLength(); aoqi@0: case vmIntrinsics::_copyOf: return inline_array_copyOf(false); aoqi@0: case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); aoqi@0: case vmIntrinsics::_equalsC: return inline_array_equals(); aoqi@0: case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); aoqi@0: aoqi@0: case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); aoqi@0: aoqi@0: case vmIntrinsics::_isInstance: aoqi@0: case vmIntrinsics::_getModifiers: aoqi@0: case vmIntrinsics::_isInterface: aoqi@0: case vmIntrinsics::_isArray: aoqi@0: case vmIntrinsics::_isPrimitive: aoqi@0: case vmIntrinsics::_getSuperclass: aoqi@0: case vmIntrinsics::_getComponentType: aoqi@0: case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); aoqi@0: aoqi@0: case vmIntrinsics::_floatToRawIntBits: aoqi@0: case vmIntrinsics::_floatToIntBits: aoqi@0: case vmIntrinsics::_intBitsToFloat: aoqi@0: case vmIntrinsics::_doubleToRawLongBits: aoqi@0: case vmIntrinsics::_doubleToLongBits: aoqi@0: case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); aoqi@0: aoqi@0: case vmIntrinsics::_numberOfLeadingZeros_i: aoqi@0: case vmIntrinsics::_numberOfLeadingZeros_l: aoqi@0: case vmIntrinsics::_numberOfTrailingZeros_i: aoqi@0: case vmIntrinsics::_numberOfTrailingZeros_l: aoqi@0: case vmIntrinsics::_bitCount_i: aoqi@0: case vmIntrinsics::_bitCount_l: aoqi@0: case vmIntrinsics::_reverseBytes_i: aoqi@0: case vmIntrinsics::_reverseBytes_l: aoqi@0: case vmIntrinsics::_reverseBytes_s: aoqi@0: case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); aoqi@0: aoqi@0: case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); aoqi@0: aoqi@0: case vmIntrinsics::_Reference_get: return inline_reference_get(); aoqi@0: aoqi@0: case vmIntrinsics::_aescrypt_encryptBlock: aoqi@0: case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); aoqi@0: aoqi@0: case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: aoqi@0: case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: aoqi@0: return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); aoqi@0: aoqi@0: case vmIntrinsics::_encodeISOArray: aoqi@0: return inline_encodeISOArray(); aoqi@0: aoqi@0: case vmIntrinsics::_updateCRC32: aoqi@0: return inline_updateCRC32(); aoqi@0: case vmIntrinsics::_updateBytesCRC32: aoqi@0: return inline_updateBytesCRC32(); aoqi@0: case vmIntrinsics::_updateByteBufferCRC32: aoqi@0: return inline_updateByteBufferCRC32(); aoqi@0: aoqi@0: default: aoqi@0: // If you get here, it may be that someone has added a new intrinsic aoqi@0: // to the list in vmSymbols.hpp without implementing it here. aoqi@0: #ifndef PRODUCT aoqi@0: if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { aoqi@0: tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", aoqi@0: vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); aoqi@0: } aoqi@0: #endif aoqi@0: return false; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: Node* LibraryCallKit::try_to_predicate() { aoqi@0: if (!jvms()->has_method()) { aoqi@0: // Root JVMState has a null method. aoqi@0: assert(map()->memory()->Opcode() == Op_Parm, ""); aoqi@0: // Insert the memory aliasing node aoqi@0: set_all_memory(reset_memory()); aoqi@0: } aoqi@0: assert(merged_memory(), ""); aoqi@0: aoqi@0: switch (intrinsic_id()) { aoqi@0: case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: aoqi@0: return inline_cipherBlockChaining_AESCrypt_predicate(false); aoqi@0: case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: aoqi@0: return inline_cipherBlockChaining_AESCrypt_predicate(true); aoqi@0: aoqi@0: default: aoqi@0: // If you get here, it may be that someone has added a new intrinsic aoqi@0: // to the list in vmSymbols.hpp without implementing it here. aoqi@0: #ifndef PRODUCT aoqi@0: if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { aoqi@0: tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", aoqi@0: vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); aoqi@0: } aoqi@0: #endif aoqi@0: Node* slow_ctl = control(); aoqi@0: set_control(top()); // No fast path instrinsic aoqi@0: return slow_ctl; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: //------------------------------set_result------------------------------- aoqi@0: // Helper function for finishing intrinsics. aoqi@0: void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { aoqi@0: record_for_igvn(region); aoqi@0: set_control(_gvn.transform(region)); aoqi@0: set_result( _gvn.transform(value)); aoqi@0: assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); aoqi@0: } aoqi@0: aoqi@0: //------------------------------generate_guard--------------------------- aoqi@0: // Helper function for generating guarded fast-slow graph structures. aoqi@0: // The given 'test', if true, guards a slow path. If the test fails aoqi@0: // then a fast path can be taken. (We generally hope it fails.) aoqi@0: // In all cases, GraphKit::control() is updated to the fast path. aoqi@0: // The returned value represents the control for the slow path. aoqi@0: // The return value is never 'top'; it is either a valid control aoqi@0: // or NULL if it is obvious that the slow path can never be taken. aoqi@0: // Also, if region and the slow control are not NULL, the slow edge aoqi@0: // is appended to the region. aoqi@0: Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { aoqi@0: if (stopped()) { aoqi@0: // Already short circuited. aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: // Build an if node and its projections. aoqi@0: // If test is true we take the slow path, which we assume is uncommon. aoqi@0: if (_gvn.type(test) == TypeInt::ZERO) { aoqi@0: // The slow branch is never taken. No need to build this guard. aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); aoqi@0: aoqi@0: Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff)); aoqi@0: if (if_slow == top()) { aoqi@0: // The slow branch is never taken. No need to build this guard. aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: if (region != NULL) aoqi@0: region->add_req(if_slow); aoqi@0: aoqi@0: Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff)); aoqi@0: set_control(if_fast); aoqi@0: aoqi@0: return if_slow; aoqi@0: } aoqi@0: aoqi@0: inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { aoqi@0: return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); aoqi@0: } aoqi@0: inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { aoqi@0: return generate_guard(test, region, PROB_FAIR); aoqi@0: } aoqi@0: aoqi@0: inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, aoqi@0: Node* *pos_index) { aoqi@0: if (stopped()) aoqi@0: return NULL; // already stopped aoqi@0: if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] aoqi@0: return NULL; // index is already adequately typed aoqi@0: Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0))); aoqi@0: Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); aoqi@0: Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); aoqi@0: if (is_neg != NULL && pos_index != NULL) { aoqi@0: // Emulate effect of Parse::adjust_map_after_if. aoqi@0: Node* ccast = new (C) CastIINode(index, TypeInt::POS); aoqi@0: ccast->set_req(0, control()); aoqi@0: (*pos_index) = _gvn.transform(ccast); aoqi@0: } aoqi@0: return is_neg; aoqi@0: } aoqi@0: aoqi@0: inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, aoqi@0: Node* *pos_index) { aoqi@0: if (stopped()) aoqi@0: return NULL; // already stopped aoqi@0: if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] aoqi@0: return NULL; // index is already adequately typed aoqi@0: Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0))); aoqi@0: BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); aoqi@0: Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq)); aoqi@0: Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); aoqi@0: if (is_notp != NULL && pos_index != NULL) { aoqi@0: // Emulate effect of Parse::adjust_map_after_if. aoqi@0: Node* ccast = new (C) CastIINode(index, TypeInt::POS1); aoqi@0: ccast->set_req(0, control()); aoqi@0: (*pos_index) = _gvn.transform(ccast); aoqi@0: } aoqi@0: return is_notp; aoqi@0: } aoqi@0: aoqi@0: // Make sure that 'position' is a valid limit index, in [0..length]. aoqi@0: // There are two equivalent plans for checking this: aoqi@0: // A. (offset + copyLength) unsigned<= arrayLength aoqi@0: // B. offset <= (arrayLength - copyLength) aoqi@0: // We require that all of the values above, except for the sum and aoqi@0: // difference, are already known to be non-negative. aoqi@0: // Plan A is robust in the face of overflow, if offset and copyLength aoqi@0: // are both hugely positive. aoqi@0: // aoqi@0: // Plan B is less direct and intuitive, but it does not overflow at aoqi@0: // all, since the difference of two non-negatives is always aoqi@0: // representable. Whenever Java methods must perform the equivalent aoqi@0: // check they generally use Plan B instead of Plan A. aoqi@0: // For the moment we use Plan A. aoqi@0: inline Node* LibraryCallKit::generate_limit_guard(Node* offset, aoqi@0: Node* subseq_length, aoqi@0: Node* array_length, aoqi@0: RegionNode* region) { aoqi@0: if (stopped()) aoqi@0: return NULL; // already stopped aoqi@0: bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; aoqi@0: if (zero_offset && subseq_length->eqv_uncast(array_length)) aoqi@0: return NULL; // common case of whole-array copy aoqi@0: Node* last = subseq_length; aoqi@0: if (!zero_offset) // last += offset aoqi@0: last = _gvn.transform(new (C) AddINode(last, offset)); aoqi@0: Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last)); aoqi@0: Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); aoqi@0: Node* is_over = generate_guard(bol_lt, region, PROB_MIN); aoqi@0: return is_over; aoqi@0: } aoqi@0: aoqi@0: aoqi@0: //--------------------------generate_current_thread-------------------- aoqi@0: Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { aoqi@0: ciKlass* thread_klass = env()->Thread_klass(); aoqi@0: const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); aoqi@0: Node* thread = _gvn.transform(new (C) ThreadLocalNode()); aoqi@0: Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); aoqi@0: Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); aoqi@0: tls_output = thread; aoqi@0: return threadObj; aoqi@0: } aoqi@0: aoqi@0: aoqi@0: //------------------------------make_string_method_node------------------------ aoqi@0: // Helper method for String intrinsic functions. This version is called aoqi@0: // with str1 and str2 pointing to String object nodes. aoqi@0: // aoqi@0: Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { aoqi@0: Node* no_ctrl = NULL; aoqi@0: aoqi@0: // Get start addr of string aoqi@0: Node* str1_value = load_String_value(no_ctrl, str1); aoqi@0: Node* str1_offset = load_String_offset(no_ctrl, str1); aoqi@0: Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); aoqi@0: aoqi@0: // Get length of string 1 aoqi@0: Node* str1_len = load_String_length(no_ctrl, str1); aoqi@0: aoqi@0: Node* str2_value = load_String_value(no_ctrl, str2); aoqi@0: Node* str2_offset = load_String_offset(no_ctrl, str2); aoqi@0: Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); aoqi@0: aoqi@0: Node* str2_len = NULL; aoqi@0: Node* result = NULL; aoqi@0: aoqi@0: switch (opcode) { aoqi@0: case Op_StrIndexOf: aoqi@0: // Get length of string 2 aoqi@0: str2_len = load_String_length(no_ctrl, str2); aoqi@0: aoqi@0: result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), aoqi@0: str1_start, str1_len, str2_start, str2_len); aoqi@0: break; aoqi@0: case Op_StrComp: aoqi@0: // Get length of string 2 aoqi@0: str2_len = load_String_length(no_ctrl, str2); aoqi@0: aoqi@0: result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), aoqi@0: str1_start, str1_len, str2_start, str2_len); aoqi@0: break; aoqi@0: case Op_StrEquals: aoqi@0: result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), aoqi@0: str1_start, str2_start, str1_len); aoqi@0: break; aoqi@0: default: aoqi@0: ShouldNotReachHere(); aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: // All these intrinsics have checks. aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: aoqi@0: return _gvn.transform(result); aoqi@0: } aoqi@0: aoqi@0: // Helper method for String intrinsic functions. This version is called aoqi@0: // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing aoqi@0: // to Int nodes containing the lenghts of str1 and str2. aoqi@0: // aoqi@0: Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { aoqi@0: Node* result = NULL; aoqi@0: switch (opcode) { aoqi@0: case Op_StrIndexOf: aoqi@0: result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), aoqi@0: str1_start, cnt1, str2_start, cnt2); aoqi@0: break; aoqi@0: case Op_StrComp: aoqi@0: result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), aoqi@0: str1_start, cnt1, str2_start, cnt2); aoqi@0: break; aoqi@0: case Op_StrEquals: aoqi@0: result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), aoqi@0: str1_start, str2_start, cnt1); aoqi@0: break; aoqi@0: default: aoqi@0: ShouldNotReachHere(); aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: // All these intrinsics have checks. aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: aoqi@0: return _gvn.transform(result); aoqi@0: } aoqi@0: aoqi@0: //------------------------------inline_string_compareTo------------------------ aoqi@0: // public int java.lang.String.compareTo(String anotherString); aoqi@0: bool LibraryCallKit::inline_string_compareTo() { aoqi@0: Node* receiver = null_check(argument(0)); aoqi@0: Node* arg = null_check(argument(1)); aoqi@0: if (stopped()) { aoqi@0: return true; aoqi@0: } aoqi@0: set_result(make_string_method_node(Op_StrComp, receiver, arg)); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------------inline_string_equals------------------------ aoqi@0: bool LibraryCallKit::inline_string_equals() { aoqi@0: Node* receiver = null_check_receiver(); aoqi@0: // NOTE: Do not null check argument for String.equals() because spec aoqi@0: // allows to specify NULL as argument. aoqi@0: Node* argument = this->argument(1); aoqi@0: if (stopped()) { aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: // paths (plus control) merge aoqi@0: RegionNode* region = new (C) RegionNode(5); aoqi@0: Node* phi = new (C) PhiNode(region, TypeInt::BOOL); aoqi@0: aoqi@0: // does source == target string? aoqi@0: Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument)); aoqi@0: Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); aoqi@0: aoqi@0: Node* if_eq = generate_slow_guard(bol, NULL); aoqi@0: if (if_eq != NULL) { aoqi@0: // receiver == argument aoqi@0: phi->init_req(2, intcon(1)); aoqi@0: region->init_req(2, if_eq); aoqi@0: } aoqi@0: aoqi@0: // get String klass for instanceOf aoqi@0: ciInstanceKlass* klass = env()->String_klass(); aoqi@0: aoqi@0: if (!stopped()) { aoqi@0: Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); aoqi@0: Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1))); aoqi@0: Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); aoqi@0: aoqi@0: Node* inst_false = generate_guard(bol, NULL, PROB_MIN); aoqi@0: //instanceOf == true, fallthrough aoqi@0: aoqi@0: if (inst_false != NULL) { aoqi@0: phi->init_req(3, intcon(0)); aoqi@0: region->init_req(3, inst_false); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: if (!stopped()) { aoqi@0: const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); aoqi@0: aoqi@0: // Properly cast the argument to String aoqi@0: argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type)); aoqi@0: // This path is taken only when argument's type is String:NotNull. aoqi@0: argument = cast_not_null(argument, false); aoqi@0: aoqi@0: Node* no_ctrl = NULL; aoqi@0: aoqi@0: // Get start addr of receiver aoqi@0: Node* receiver_val = load_String_value(no_ctrl, receiver); aoqi@0: Node* receiver_offset = load_String_offset(no_ctrl, receiver); aoqi@0: Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); aoqi@0: aoqi@0: // Get length of receiver aoqi@0: Node* receiver_cnt = load_String_length(no_ctrl, receiver); aoqi@0: aoqi@0: // Get start addr of argument aoqi@0: Node* argument_val = load_String_value(no_ctrl, argument); aoqi@0: Node* argument_offset = load_String_offset(no_ctrl, argument); aoqi@0: Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); aoqi@0: aoqi@0: // Get length of argument aoqi@0: Node* argument_cnt = load_String_length(no_ctrl, argument); aoqi@0: aoqi@0: // Check for receiver count != argument count aoqi@0: Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt)); aoqi@0: Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne)); aoqi@0: Node* if_ne = generate_slow_guard(bol, NULL); aoqi@0: if (if_ne != NULL) { aoqi@0: phi->init_req(4, intcon(0)); aoqi@0: region->init_req(4, if_ne); aoqi@0: } aoqi@0: aoqi@0: // Check for count == 0 is done by assembler code for StrEquals. aoqi@0: aoqi@0: if (!stopped()) { aoqi@0: Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); aoqi@0: phi->init_req(1, equals); aoqi@0: region->init_req(1, control()); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // post merge aoqi@0: set_control(_gvn.transform(region)); aoqi@0: record_for_igvn(region); aoqi@0: aoqi@0: set_result(_gvn.transform(phi)); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------------inline_array_equals---------------------------- aoqi@0: bool LibraryCallKit::inline_array_equals() { aoqi@0: Node* arg1 = argument(0); aoqi@0: Node* arg2 = argument(1); aoqi@0: set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: // Java version of String.indexOf(constant string) aoqi@0: // class StringDecl { aoqi@0: // StringDecl(char[] ca) { aoqi@0: // offset = 0; aoqi@0: // count = ca.length; aoqi@0: // value = ca; aoqi@0: // } aoqi@0: // int offset; aoqi@0: // int count; aoqi@0: // char[] value; aoqi@0: // } aoqi@0: // aoqi@0: // static int string_indexOf_J(StringDecl string_object, char[] target_object, aoqi@0: // int targetOffset, int cache_i, int md2) { aoqi@0: // int cache = cache_i; aoqi@0: // int sourceOffset = string_object.offset; aoqi@0: // int sourceCount = string_object.count; aoqi@0: // int targetCount = target_object.length; aoqi@0: // aoqi@0: // int targetCountLess1 = targetCount - 1; aoqi@0: // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; aoqi@0: // aoqi@0: // char[] source = string_object.value; aoqi@0: // char[] target = target_object; aoqi@0: // int lastChar = target[targetCountLess1]; aoqi@0: // aoqi@0: // outer_loop: aoqi@0: // for (int i = sourceOffset; i < sourceEnd; ) { aoqi@0: // int src = source[i + targetCountLess1]; aoqi@0: // if (src == lastChar) { aoqi@0: // // With random strings and a 4-character alphabet, aoqi@0: // // reverse matching at this point sets up 0.8% fewer aoqi@0: // // frames, but (paradoxically) makes 0.3% more probes. aoqi@0: // // Since those probes are nearer the lastChar probe, aoqi@0: // // there is may be a net D$ win with reverse matching. aoqi@0: // // But, reversing loop inhibits unroll of inner loop aoqi@0: // // for unknown reason. So, does running outer loop from aoqi@0: // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) aoqi@0: // for (int j = 0; j < targetCountLess1; j++) { aoqi@0: // if (target[targetOffset + j] != source[i+j]) { aoqi@0: // if ((cache & (1 << source[i+j])) == 0) { aoqi@0: // if (md2 < j+1) { aoqi@0: // i += j+1; aoqi@0: // continue outer_loop; aoqi@0: // } aoqi@0: // } aoqi@0: // i += md2; aoqi@0: // continue outer_loop; aoqi@0: // } aoqi@0: // } aoqi@0: // return i - sourceOffset; aoqi@0: // } aoqi@0: // if ((cache & (1 << src)) == 0) { aoqi@0: // i += targetCountLess1; aoqi@0: // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. aoqi@0: // i++; aoqi@0: // } aoqi@0: // return -1; aoqi@0: // } aoqi@0: aoqi@0: //------------------------------string_indexOf------------------------ aoqi@0: Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, aoqi@0: jint cache_i, jint md2_i) { aoqi@0: aoqi@0: Node* no_ctrl = NULL; aoqi@0: float likely = PROB_LIKELY(0.9); aoqi@0: float unlikely = PROB_UNLIKELY(0.9); aoqi@0: aoqi@0: const int nargs = 0; // no arguments to push back for uncommon trap in predicate aoqi@0: aoqi@0: Node* source = load_String_value(no_ctrl, string_object); aoqi@0: Node* sourceOffset = load_String_offset(no_ctrl, string_object); aoqi@0: Node* sourceCount = load_String_length(no_ctrl, string_object); aoqi@0: aoqi@0: Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); aoqi@0: jint target_length = target_array->length(); aoqi@0: const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); aoqi@0: const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); aoqi@0: aoqi@0: // String.value field is known to be @Stable. aoqi@0: if (UseImplicitStableValues) { aoqi@0: target = cast_array_to_stable(target, target_type); aoqi@0: } aoqi@0: aoqi@0: IdealKit kit(this, false, true); aoqi@0: #define __ kit. aoqi@0: Node* zero = __ ConI(0); aoqi@0: Node* one = __ ConI(1); aoqi@0: Node* cache = __ ConI(cache_i); aoqi@0: Node* md2 = __ ConI(md2_i); aoqi@0: Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); aoqi@0: Node* targetCount = __ ConI(target_length); aoqi@0: Node* targetCountLess1 = __ ConI(target_length - 1); aoqi@0: Node* targetOffset = __ ConI(targetOffset_i); aoqi@0: Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); aoqi@0: aoqi@0: IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); aoqi@0: Node* outer_loop = __ make_label(2 /* goto */); aoqi@0: Node* return_ = __ make_label(1); aoqi@0: aoqi@0: __ set(rtn,__ ConI(-1)); aoqi@0: __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { aoqi@0: Node* i2 = __ AddI(__ value(i), targetCountLess1); aoqi@0: // pin to prohibit loading of "next iteration" value which may SEGV (rare) aoqi@0: Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); aoqi@0: __ if_then(src, BoolTest::eq, lastChar, unlikely); { aoqi@0: __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { aoqi@0: Node* tpj = __ AddI(targetOffset, __ value(j)); aoqi@0: Node* targ = load_array_element(no_ctrl, target, tpj, target_type); aoqi@0: Node* ipj = __ AddI(__ value(i), __ value(j)); aoqi@0: Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); aoqi@0: __ if_then(targ, BoolTest::ne, src2); { aoqi@0: __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { aoqi@0: __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { aoqi@0: __ increment(i, __ AddI(__ value(j), one)); aoqi@0: __ goto_(outer_loop); aoqi@0: } __ end_if(); __ dead(j); aoqi@0: }__ end_if(); __ dead(j); aoqi@0: __ increment(i, md2); aoqi@0: __ goto_(outer_loop); aoqi@0: }__ end_if(); aoqi@0: __ increment(j, one); aoqi@0: }__ end_loop(); __ dead(j); aoqi@0: __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); aoqi@0: __ goto_(return_); aoqi@0: }__ end_if(); aoqi@0: __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { aoqi@0: __ increment(i, targetCountLess1); aoqi@0: }__ end_if(); aoqi@0: __ increment(i, one); aoqi@0: __ bind(outer_loop); aoqi@0: }__ end_loop(); __ dead(i); aoqi@0: __ bind(return_); aoqi@0: aoqi@0: // Final sync IdealKit and GraphKit. aoqi@0: final_sync(kit); aoqi@0: Node* result = __ value(rtn); aoqi@0: #undef __ aoqi@0: C->set_has_loops(true); aoqi@0: return result; aoqi@0: } aoqi@0: aoqi@0: //------------------------------inline_string_indexOf------------------------ aoqi@0: bool LibraryCallKit::inline_string_indexOf() { aoqi@0: Node* receiver = argument(0); aoqi@0: Node* arg = argument(1); aoqi@0: aoqi@0: Node* result; aoqi@0: // Disable the use of pcmpestri until it can be guaranteed that aoqi@0: // the load doesn't cross into the uncommited space. aoqi@0: if (Matcher::has_match_rule(Op_StrIndexOf) && aoqi@0: UseSSE42Intrinsics) { aoqi@0: // Generate SSE4.2 version of indexOf aoqi@0: // We currently only have match rules that use SSE4.2 aoqi@0: aoqi@0: receiver = null_check(receiver); aoqi@0: arg = null_check(arg); aoqi@0: if (stopped()) { aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: ciInstanceKlass* str_klass = env()->String_klass(); aoqi@0: const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); aoqi@0: aoqi@0: // Make the merge point aoqi@0: RegionNode* result_rgn = new (C) RegionNode(4); aoqi@0: Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT); aoqi@0: Node* no_ctrl = NULL; aoqi@0: aoqi@0: // Get start addr of source string aoqi@0: Node* source = load_String_value(no_ctrl, receiver); aoqi@0: Node* source_offset = load_String_offset(no_ctrl, receiver); aoqi@0: Node* source_start = array_element_address(source, source_offset, T_CHAR); aoqi@0: aoqi@0: // Get length of source string aoqi@0: Node* source_cnt = load_String_length(no_ctrl, receiver); aoqi@0: aoqi@0: // Get start addr of substring aoqi@0: Node* substr = load_String_value(no_ctrl, arg); aoqi@0: Node* substr_offset = load_String_offset(no_ctrl, arg); aoqi@0: Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); aoqi@0: aoqi@0: // Get length of source string aoqi@0: Node* substr_cnt = load_String_length(no_ctrl, arg); aoqi@0: aoqi@0: // Check for substr count > string count aoqi@0: Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt)); aoqi@0: Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt)); aoqi@0: Node* if_gt = generate_slow_guard(bol, NULL); aoqi@0: if (if_gt != NULL) { aoqi@0: result_phi->init_req(2, intcon(-1)); aoqi@0: result_rgn->init_req(2, if_gt); aoqi@0: } aoqi@0: aoqi@0: if (!stopped()) { aoqi@0: // Check for substr count == 0 aoqi@0: cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0))); aoqi@0: bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); aoqi@0: Node* if_zero = generate_slow_guard(bol, NULL); aoqi@0: if (if_zero != NULL) { aoqi@0: result_phi->init_req(3, intcon(0)); aoqi@0: result_rgn->init_req(3, if_zero); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: if (!stopped()) { aoqi@0: result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); aoqi@0: result_phi->init_req(1, result); aoqi@0: result_rgn->init_req(1, control()); aoqi@0: } aoqi@0: set_control(_gvn.transform(result_rgn)); aoqi@0: record_for_igvn(result_rgn); aoqi@0: result = _gvn.transform(result_phi); aoqi@0: aoqi@0: } else { // Use LibraryCallKit::string_indexOf aoqi@0: // don't intrinsify if argument isn't a constant string. aoqi@0: if (!arg->is_Con()) { aoqi@0: return false; aoqi@0: } aoqi@0: const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); aoqi@0: if (str_type == NULL) { aoqi@0: return false; aoqi@0: } aoqi@0: ciInstanceKlass* klass = env()->String_klass(); aoqi@0: ciObject* str_const = str_type->const_oop(); aoqi@0: if (str_const == NULL || str_const->klass() != klass) { aoqi@0: return false; aoqi@0: } aoqi@0: ciInstance* str = str_const->as_instance(); aoqi@0: assert(str != NULL, "must be instance"); aoqi@0: aoqi@0: ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); aoqi@0: ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array aoqi@0: aoqi@0: int o; aoqi@0: int c; aoqi@0: if (java_lang_String::has_offset_field()) { aoqi@0: o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); aoqi@0: c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); aoqi@0: } else { aoqi@0: o = 0; aoqi@0: c = pat->length(); aoqi@0: } aoqi@0: aoqi@0: // constant strings have no offset and count == length which aoqi@0: // simplifies the resulting code somewhat so lets optimize for that. aoqi@0: if (o != 0 || c != pat->length()) { aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: receiver = null_check(receiver, T_OBJECT); aoqi@0: // NOTE: No null check on the argument is needed since it's a constant String oop. aoqi@0: if (stopped()) { aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: // The null string as a pattern always returns 0 (match at beginning of string) aoqi@0: if (c == 0) { aoqi@0: set_result(intcon(0)); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: // Generate default indexOf aoqi@0: jchar lastChar = pat->char_at(o + (c - 1)); aoqi@0: int cache = 0; aoqi@0: int i; aoqi@0: for (i = 0; i < c - 1; i++) { aoqi@0: assert(i < pat->length(), "out of range"); aoqi@0: cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); aoqi@0: } aoqi@0: aoqi@0: int md2 = c; aoqi@0: for (i = 0; i < c - 1; i++) { aoqi@0: assert(i < pat->length(), "out of range"); aoqi@0: if (pat->char_at(o + i) == lastChar) { aoqi@0: md2 = (c - 1) - i; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: result = string_indexOf(receiver, pat, o, cache, md2); aoqi@0: } aoqi@0: set_result(result); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //--------------------------round_double_node-------------------------------- aoqi@0: // Round a double node if necessary. aoqi@0: Node* LibraryCallKit::round_double_node(Node* n) { aoqi@0: if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) aoqi@0: n = _gvn.transform(new (C) RoundDoubleNode(0, n)); aoqi@0: return n; aoqi@0: } aoqi@0: aoqi@0: //------------------------------inline_math----------------------------------- aoqi@0: // public static double Math.abs(double) aoqi@0: // public static double Math.sqrt(double) aoqi@0: // public static double Math.log(double) aoqi@0: // public static double Math.log10(double) aoqi@0: bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { aoqi@0: Node* arg = round_double_node(argument(0)); aoqi@0: Node* n; aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break; aoqi@0: case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break; aoqi@0: case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break; aoqi@0: case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break; aoqi@0: default: fatal_unexpected_iid(id); break; aoqi@0: } aoqi@0: set_result(_gvn.transform(n)); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------------inline_trig---------------------------------- aoqi@0: // Inline sin/cos/tan instructions, if possible. If rounding is required, do aoqi@0: // argument reduction which will turn into a fast/slow diamond. aoqi@0: bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { aoqi@0: Node* arg = round_double_node(argument(0)); aoqi@0: Node* n = NULL; aoqi@0: aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break; aoqi@0: case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break; aoqi@0: case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break; aoqi@0: default: fatal_unexpected_iid(id); break; aoqi@0: } aoqi@0: n = _gvn.transform(n); aoqi@0: aoqi@0: // Rounding required? Check for argument reduction! aoqi@0: if (Matcher::strict_fp_requires_explicit_rounding) { aoqi@0: static const double pi_4 = 0.7853981633974483; aoqi@0: static const double neg_pi_4 = -0.7853981633974483; aoqi@0: // pi/2 in 80-bit extended precision aoqi@0: // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; aoqi@0: // -pi/2 in 80-bit extended precision aoqi@0: // 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}; aoqi@0: // Cutoff value for using this argument reduction technique aoqi@0: //static const double pi_2_minus_epsilon = 1.564660403643354; aoqi@0: //static const double neg_pi_2_plus_epsilon = -1.564660403643354; aoqi@0: aoqi@0: // Pseudocode for sin: aoqi@0: // if (x <= Math.PI / 4.0) { aoqi@0: // if (x >= -Math.PI / 4.0) return fsin(x); aoqi@0: // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); aoqi@0: // } else { aoqi@0: // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); aoqi@0: // } aoqi@0: // return StrictMath.sin(x); aoqi@0: aoqi@0: // Pseudocode for cos: aoqi@0: // if (x <= Math.PI / 4.0) { aoqi@0: // if (x >= -Math.PI / 4.0) return fcos(x); aoqi@0: // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); aoqi@0: // } else { aoqi@0: // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); aoqi@0: // } aoqi@0: // return StrictMath.cos(x); aoqi@0: aoqi@0: // Actually, sticking in an 80-bit Intel value into C2 will be tough; it aoqi@0: // requires a special machine instruction to load it. Instead we'll try aoqi@0: // the 'easy' case. If we really need the extra range +/- PI/2 we'll aoqi@0: // probably do the math inside the SIN encoding. aoqi@0: aoqi@0: // Make the merge point aoqi@0: RegionNode* r = new (C) RegionNode(3); aoqi@0: Node* phi = new (C) PhiNode(r, Type::DOUBLE); aoqi@0: aoqi@0: // Flatten arg so we need only 1 test aoqi@0: Node *abs = _gvn.transform(new (C) AbsDNode(arg)); aoqi@0: // Node for PI/4 constant aoqi@0: Node *pi4 = makecon(TypeD::make(pi_4)); aoqi@0: // Check PI/4 : abs(arg) aoqi@0: Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs)); aoqi@0: // Check: If PI/4 < abs(arg) then go slow aoqi@0: Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt )); aoqi@0: // Branch either way aoqi@0: IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); aoqi@0: set_control(opt_iff(r,iff)); aoqi@0: aoqi@0: // Set fast path result aoqi@0: phi->init_req(2, n); aoqi@0: aoqi@0: // Slow path - non-blocking leaf call aoqi@0: Node* call = NULL; aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_dsin: aoqi@0: call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), aoqi@0: CAST_FROM_FN_PTR(address, SharedRuntime::dsin), aoqi@0: "Sin", NULL, arg, top()); aoqi@0: break; aoqi@0: case vmIntrinsics::_dcos: aoqi@0: call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), aoqi@0: CAST_FROM_FN_PTR(address, SharedRuntime::dcos), aoqi@0: "Cos", NULL, arg, top()); aoqi@0: break; aoqi@0: case vmIntrinsics::_dtan: aoqi@0: call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), aoqi@0: CAST_FROM_FN_PTR(address, SharedRuntime::dtan), aoqi@0: "Tan", NULL, arg, top()); aoqi@0: break; aoqi@0: } aoqi@0: assert(control()->in(0) == call, ""); aoqi@0: Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); aoqi@0: r->init_req(1, control()); aoqi@0: phi->init_req(1, slow_result); aoqi@0: aoqi@0: // Post-merge aoqi@0: set_control(_gvn.transform(r)); aoqi@0: record_for_igvn(r); aoqi@0: n = _gvn.transform(phi); aoqi@0: aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: } aoqi@0: set_result(n); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { aoqi@0: //------------------- aoqi@0: //result=(result.isNaN())? funcAddr():result; aoqi@0: // Check: If isNaN() by checking result!=result? then either trap aoqi@0: // or go to runtime aoqi@0: Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result)); aoqi@0: // Build the boolean node aoqi@0: Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq)); aoqi@0: aoqi@0: if (!too_many_traps(Deoptimization::Reason_intrinsic)) { aoqi@0: { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); aoqi@0: // The pow or exp intrinsic returned a NaN, which requires a call aoqi@0: // to the runtime. Recompile with the runtime call. aoqi@0: uncommon_trap(Deoptimization::Reason_intrinsic, aoqi@0: Deoptimization::Action_make_not_entrant); aoqi@0: } aoqi@0: return result; aoqi@0: } else { aoqi@0: // If this inlining ever returned NaN in the past, we compile a call aoqi@0: // to the runtime to properly handle corner cases aoqi@0: aoqi@0: IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); aoqi@0: Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff)); aoqi@0: Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff)); aoqi@0: aoqi@0: if (!if_slow->is_top()) { aoqi@0: RegionNode* result_region = new (C) RegionNode(3); aoqi@0: PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE); aoqi@0: aoqi@0: result_region->init_req(1, if_fast); aoqi@0: result_val->init_req(1, result); aoqi@0: aoqi@0: set_control(if_slow); aoqi@0: aoqi@0: const TypePtr* no_memory_effects = NULL; aoqi@0: Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, aoqi@0: no_memory_effects, aoqi@0: x, top(), y, y ? top() : NULL); aoqi@0: Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0)); aoqi@0: #ifdef ASSERT aoqi@0: Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1)); aoqi@0: assert(value_top == top(), "second value must be top"); aoqi@0: #endif aoqi@0: aoqi@0: result_region->init_req(2, control()); aoqi@0: result_val->init_req(2, value); aoqi@0: set_control(_gvn.transform(result_region)); aoqi@0: return _gvn.transform(result_val); aoqi@0: } else { aoqi@0: return result; aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: //------------------------------inline_exp------------------------------------- aoqi@0: // Inline exp instructions, if possible. The Intel hardware only misses aoqi@0: // really odd corner cases (+/- Infinity). Just uncommon-trap them. aoqi@0: bool LibraryCallKit::inline_exp() { aoqi@0: Node* arg = round_double_node(argument(0)); aoqi@0: Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg)); aoqi@0: aoqi@0: n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); aoqi@0: set_result(n); aoqi@0: aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------------inline_pow------------------------------------- aoqi@0: // Inline power instructions, if possible. aoqi@0: bool LibraryCallKit::inline_pow() { aoqi@0: // Pseudocode for pow aoqi@0: // if (y == 2) { aoqi@0: // return x * x; aoqi@0: // } else { aoqi@0: // if (x <= 0.0) { aoqi@0: // long longy = (long)y; aoqi@0: // if ((double)longy == y) { // if y is long aoqi@0: // if (y + 1 == y) longy = 0; // huge number: even aoqi@0: // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); aoqi@0: // } else { aoqi@0: // result = NaN; aoqi@0: // } aoqi@0: // } else { aoqi@0: // result = DPow(x,y); aoqi@0: // } aoqi@0: // if (result != result)? { aoqi@0: // result = uncommon_trap() or runtime_call(); aoqi@0: // } aoqi@0: // return result; aoqi@0: // } aoqi@0: aoqi@0: Node* x = round_double_node(argument(0)); aoqi@0: Node* y = round_double_node(argument(2)); aoqi@0: aoqi@0: Node* result = NULL; aoqi@0: aoqi@0: Node* const_two_node = makecon(TypeD::make(2.0)); aoqi@0: Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node)); aoqi@0: Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq)); aoqi@0: IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); aoqi@0: Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node)); aoqi@0: Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node)); aoqi@0: aoqi@0: RegionNode* region_node = new (C) RegionNode(3); aoqi@0: region_node->init_req(1, if_true); aoqi@0: aoqi@0: Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE); aoqi@0: // special case for x^y where y == 2, we can convert it to x * x aoqi@0: phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x))); aoqi@0: aoqi@0: // set control to if_false since we will now process the false branch aoqi@0: set_control(if_false); aoqi@0: aoqi@0: if (!too_many_traps(Deoptimization::Reason_intrinsic)) { aoqi@0: // Short form: skip the fancy tests and just check for NaN result. aoqi@0: result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); aoqi@0: } else { aoqi@0: // If this inlining ever returned NaN in the past, include all aoqi@0: // checks + call to the runtime. aoqi@0: aoqi@0: // Set the merge point for If node with condition of (x <= 0.0) aoqi@0: // There are four possible paths to region node and phi node aoqi@0: RegionNode *r = new (C) RegionNode(4); aoqi@0: Node *phi = new (C) PhiNode(r, Type::DOUBLE); aoqi@0: aoqi@0: // Build the first if node: if (x <= 0.0) aoqi@0: // Node for 0 constant aoqi@0: Node *zeronode = makecon(TypeD::ZERO); aoqi@0: // Check x:0 aoqi@0: Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode)); aoqi@0: // Check: If (x<=0) then go complex path aoqi@0: Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le )); aoqi@0: // Branch either way aoqi@0: IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); aoqi@0: // Fast path taken; set region slot 3 aoqi@0: Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1)); aoqi@0: r->init_req(3,fast_taken); // Capture fast-control aoqi@0: aoqi@0: // Fast path not-taken, i.e. slow path aoqi@0: Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1)); aoqi@0: aoqi@0: // Set fast path result aoqi@0: Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); aoqi@0: phi->init_req(3, fast_result); aoqi@0: aoqi@0: // Complex path aoqi@0: // Build the second if node (if y is long) aoqi@0: // Node for (long)y aoqi@0: Node *longy = _gvn.transform(new (C) ConvD2LNode(y)); aoqi@0: // Node for (double)((long) y) aoqi@0: Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy)); aoqi@0: // Check (double)((long) y) : y aoqi@0: Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y)); aoqi@0: // Check if (y isn't long) then go to slow path aoqi@0: aoqi@0: Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne )); aoqi@0: // Branch either way aoqi@0: IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); aoqi@0: Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2)); aoqi@0: aoqi@0: Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2)); aoqi@0: aoqi@0: // Calculate DPow(abs(x), y)*(1 & (long)y) aoqi@0: // Node for constant 1 aoqi@0: Node *conone = longcon(1); aoqi@0: // 1& (long)y aoqi@0: Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy)); aoqi@0: aoqi@0: // A huge number is always even. Detect a huge number by checking aoqi@0: // if y + 1 == y and set integer to be tested for parity to 0. aoqi@0: // Required for corner case: aoqi@0: // (long)9.223372036854776E18 = max_jlong aoqi@0: // (double)(long)9.223372036854776E18 = 9.223372036854776E18 aoqi@0: // max_jlong is odd but 9.223372036854776E18 is even aoqi@0: Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1)))); aoqi@0: Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y)); aoqi@0: Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq )); aoqi@0: Node* correctedsign = NULL; aoqi@0: if (ConditionalMoveLimit != 0) { aoqi@0: correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); aoqi@0: } else { aoqi@0: IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); aoqi@0: RegionNode *r = new (C) RegionNode(3); aoqi@0: Node *phi = new (C) PhiNode(r, TypeLong::LONG); aoqi@0: r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1))); aoqi@0: r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1))); aoqi@0: phi->init_req(1, signnode); aoqi@0: phi->init_req(2, longcon(0)); aoqi@0: correctedsign = _gvn.transform(phi); aoqi@0: ylong_path = _gvn.transform(r); aoqi@0: record_for_igvn(r); aoqi@0: } aoqi@0: aoqi@0: // zero node aoqi@0: Node *conzero = longcon(0); aoqi@0: // Check (1&(long)y)==0? aoqi@0: Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero)); aoqi@0: // Check if (1&(long)y)!=0?, if so the result is negative aoqi@0: Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne )); aoqi@0: // abs(x) aoqi@0: Node *absx=_gvn.transform(new (C) AbsDNode(x)); aoqi@0: // abs(x)^y aoqi@0: Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y)); aoqi@0: // -abs(x)^y aoqi@0: Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy)); aoqi@0: // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) aoqi@0: Node *signresult = NULL; aoqi@0: if (ConditionalMoveLimit != 0) { aoqi@0: signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); aoqi@0: } else { aoqi@0: IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); aoqi@0: RegionNode *r = new (C) RegionNode(3); aoqi@0: Node *phi = new (C) PhiNode(r, Type::DOUBLE); aoqi@0: r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven))); aoqi@0: r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven))); aoqi@0: phi->init_req(1, absxpowy); aoqi@0: phi->init_req(2, negabsxpowy); aoqi@0: signresult = _gvn.transform(phi); aoqi@0: ylong_path = _gvn.transform(r); aoqi@0: record_for_igvn(r); aoqi@0: } aoqi@0: // Set complex path fast result aoqi@0: r->init_req(2, ylong_path); aoqi@0: phi->init_req(2, signresult); aoqi@0: aoqi@0: static const jlong nan_bits = CONST64(0x7ff8000000000000); aoqi@0: Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN aoqi@0: r->init_req(1,slow_path); aoqi@0: phi->init_req(1,slow_result); aoqi@0: aoqi@0: // Post merge aoqi@0: set_control(_gvn.transform(r)); aoqi@0: record_for_igvn(r); aoqi@0: result = _gvn.transform(phi); aoqi@0: } aoqi@0: aoqi@0: result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); aoqi@0: aoqi@0: // control from finish_pow_exp is now input to the region node aoqi@0: region_node->set_req(2, control()); aoqi@0: // the result from finish_pow_exp is now input to the phi node aoqi@0: phi_node->init_req(2, result); aoqi@0: set_control(_gvn.transform(region_node)); aoqi@0: record_for_igvn(region_node); aoqi@0: set_result(_gvn.transform(phi_node)); aoqi@0: aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------------runtime_math----------------------------- aoqi@0: bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { aoqi@0: assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), aoqi@0: "must be (DD)D or (D)D type"); aoqi@0: aoqi@0: // Inputs aoqi@0: Node* a = round_double_node(argument(0)); aoqi@0: Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; aoqi@0: aoqi@0: const TypePtr* no_memory_effects = NULL; aoqi@0: Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, aoqi@0: no_memory_effects, aoqi@0: a, top(), b, b ? top() : NULL); aoqi@0: Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0)); aoqi@0: #ifdef ASSERT aoqi@0: Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1)); aoqi@0: assert(value_top == top(), "second value must be top"); aoqi@0: #endif aoqi@0: aoqi@0: set_result(value); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------------inline_math_native----------------------------- aoqi@0: bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { aoqi@0: #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) aoqi@0: switch (id) { aoqi@0: // These intrinsics are not properly supported on all hardware aoqi@0: case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : aoqi@0: runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); aoqi@0: case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : aoqi@0: runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); aoqi@0: case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : aoqi@0: runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); aoqi@0: aoqi@0: case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : aoqi@0: runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); aoqi@0: case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : aoqi@0: runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); aoqi@0: aoqi@0: // These intrinsics are supported on all hardware aoqi@0: case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; aoqi@0: case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; aoqi@0: aoqi@0: case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : aoqi@0: runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); aoqi@0: case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : aoqi@0: runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); aoqi@0: #undef FN_PTR aoqi@0: aoqi@0: // These intrinsics are not yet correctly implemented aoqi@0: case vmIntrinsics::_datan2: aoqi@0: return false; aoqi@0: aoqi@0: default: aoqi@0: fatal_unexpected_iid(id); aoqi@0: return false; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: static bool is_simple_name(Node* n) { aoqi@0: return (n->req() == 1 // constant aoqi@0: || (n->is_Type() && n->as_Type()->type()->singleton()) aoqi@0: || n->is_Proj() // parameter or return value aoqi@0: || n->is_Phi() // local of some sort aoqi@0: ); aoqi@0: } aoqi@0: aoqi@0: //----------------------------inline_min_max----------------------------------- aoqi@0: bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { aoqi@0: set_result(generate_min_max(id, argument(0), argument(1))); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { aoqi@0: Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) ); aoqi@0: IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); aoqi@0: Node* fast_path = _gvn.transform( new (C) IfFalseNode(check)); aoqi@0: Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) ); aoqi@0: aoqi@0: { aoqi@0: PreserveJVMState pjvms(this); aoqi@0: PreserveReexecuteState preexecs(this); aoqi@0: jvms()->set_should_reexecute(true); aoqi@0: aoqi@0: set_control(slow_path); aoqi@0: set_i_o(i_o()); aoqi@0: aoqi@0: uncommon_trap(Deoptimization::Reason_intrinsic, aoqi@0: Deoptimization::Action_none); aoqi@0: } aoqi@0: aoqi@0: set_control(fast_path); aoqi@0: set_result(math); aoqi@0: } aoqi@0: aoqi@0: template aoqi@0: bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { aoqi@0: typedef typename OverflowOp::MathOp MathOp; aoqi@0: aoqi@0: MathOp* mathOp = new(C) MathOp(arg1, arg2); aoqi@0: Node* operation = _gvn.transform( mathOp ); aoqi@0: Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) ); aoqi@0: inline_math_mathExact(operation, ofcheck); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_math_addExactI(bool is_increment) { aoqi@0: return inline_math_overflow(argument(0), is_increment ? intcon(1) : argument(1)); aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_math_addExactL(bool is_increment) { aoqi@0: return inline_math_overflow(argument(0), is_increment ? longcon(1) : argument(2)); aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { aoqi@0: return inline_math_overflow(argument(0), is_decrement ? intcon(1) : argument(1)); aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { aoqi@0: return inline_math_overflow(argument(0), is_decrement ? longcon(1) : argument(2)); aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_math_negateExactI() { aoqi@0: return inline_math_overflow(intcon(0), argument(0)); aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_math_negateExactL() { aoqi@0: return inline_math_overflow(longcon(0), argument(0)); aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_math_multiplyExactI() { aoqi@0: return inline_math_overflow(argument(0), argument(1)); aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_math_multiplyExactL() { aoqi@0: return inline_math_overflow(argument(0), argument(2)); aoqi@0: } aoqi@0: aoqi@0: Node* aoqi@0: LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { aoqi@0: // These are the candidate return value: aoqi@0: Node* xvalue = x0; aoqi@0: Node* yvalue = y0; aoqi@0: aoqi@0: if (xvalue == yvalue) { aoqi@0: return xvalue; aoqi@0: } aoqi@0: aoqi@0: bool want_max = (id == vmIntrinsics::_max); aoqi@0: aoqi@0: const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); aoqi@0: const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); aoqi@0: if (txvalue == NULL || tyvalue == NULL) return top(); aoqi@0: // This is not really necessary, but it is consistent with a aoqi@0: // hypothetical MaxINode::Value method: aoqi@0: int widen = MAX2(txvalue->_widen, tyvalue->_widen); aoqi@0: aoqi@0: // %%% This folding logic should (ideally) be in a different place. aoqi@0: // Some should be inside IfNode, and there to be a more reliable aoqi@0: // transformation of ?: style patterns into cmoves. We also want aoqi@0: // more powerful optimizations around cmove and min/max. aoqi@0: aoqi@0: // Try to find a dominating comparison of these guys. aoqi@0: // It can simplify the index computation for Arrays.copyOf aoqi@0: // and similar uses of System.arraycopy. aoqi@0: // First, compute the normalized version of CmpI(x, y). aoqi@0: int cmp_op = Op_CmpI; aoqi@0: Node* xkey = xvalue; aoqi@0: Node* ykey = yvalue; aoqi@0: Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey)); aoqi@0: if (ideal_cmpxy->is_Cmp()) { aoqi@0: // E.g., if we have CmpI(length - offset, count), aoqi@0: // it might idealize to CmpI(length, count + offset) aoqi@0: cmp_op = ideal_cmpxy->Opcode(); aoqi@0: xkey = ideal_cmpxy->in(1); aoqi@0: ykey = ideal_cmpxy->in(2); aoqi@0: } aoqi@0: aoqi@0: // Start by locating any relevant comparisons. aoqi@0: Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; aoqi@0: Node* cmpxy = NULL; aoqi@0: Node* cmpyx = NULL; aoqi@0: for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { aoqi@0: Node* cmp = start_from->fast_out(k); aoqi@0: if (cmp->outcnt() > 0 && // must have prior uses aoqi@0: cmp->in(0) == NULL && // must be context-independent aoqi@0: cmp->Opcode() == cmp_op) { // right kind of compare aoqi@0: if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; aoqi@0: if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: const int NCMPS = 2; aoqi@0: Node* cmps[NCMPS] = { cmpxy, cmpyx }; aoqi@0: int cmpn; aoqi@0: for (cmpn = 0; cmpn < NCMPS; cmpn++) { aoqi@0: if (cmps[cmpn] != NULL) break; // find a result aoqi@0: } aoqi@0: if (cmpn < NCMPS) { aoqi@0: // Look for a dominating test that tells us the min and max. aoqi@0: int depth = 0; // Limit search depth for speed aoqi@0: Node* dom = control(); aoqi@0: for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { aoqi@0: if (++depth >= 100) break; aoqi@0: Node* ifproj = dom; aoqi@0: if (!ifproj->is_Proj()) continue; aoqi@0: Node* iff = ifproj->in(0); aoqi@0: if (!iff->is_If()) continue; aoqi@0: Node* bol = iff->in(1); aoqi@0: if (!bol->is_Bool()) continue; aoqi@0: Node* cmp = bol->in(1); aoqi@0: if (cmp == NULL) continue; aoqi@0: for (cmpn = 0; cmpn < NCMPS; cmpn++) aoqi@0: if (cmps[cmpn] == cmp) break; aoqi@0: if (cmpn == NCMPS) continue; aoqi@0: BoolTest::mask btest = bol->as_Bool()->_test._test; aoqi@0: if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); aoqi@0: if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); aoqi@0: // At this point, we know that 'x btest y' is true. aoqi@0: switch (btest) { aoqi@0: case BoolTest::eq: aoqi@0: // They are proven equal, so we can collapse the min/max. aoqi@0: // Either value is the answer. Choose the simpler. aoqi@0: if (is_simple_name(yvalue) && !is_simple_name(xvalue)) aoqi@0: return yvalue; aoqi@0: return xvalue; aoqi@0: case BoolTest::lt: // x < y aoqi@0: case BoolTest::le: // x <= y aoqi@0: return (want_max ? yvalue : xvalue); aoqi@0: case BoolTest::gt: // x > y aoqi@0: case BoolTest::ge: // x >= y aoqi@0: return (want_max ? xvalue : yvalue); aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // We failed to find a dominating test. aoqi@0: // Let's pick a test that might GVN with prior tests. aoqi@0: Node* best_bol = NULL; aoqi@0: BoolTest::mask best_btest = BoolTest::illegal; aoqi@0: for (cmpn = 0; cmpn < NCMPS; cmpn++) { aoqi@0: Node* cmp = cmps[cmpn]; aoqi@0: if (cmp == NULL) continue; aoqi@0: for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { aoqi@0: Node* bol = cmp->fast_out(j); aoqi@0: if (!bol->is_Bool()) continue; aoqi@0: BoolTest::mask btest = bol->as_Bool()->_test._test; aoqi@0: if (btest == BoolTest::eq || btest == BoolTest::ne) continue; aoqi@0: if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); aoqi@0: if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { aoqi@0: best_bol = bol->as_Bool(); aoqi@0: best_btest = btest; aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: Node* answer_if_true = NULL; aoqi@0: Node* answer_if_false = NULL; aoqi@0: switch (best_btest) { aoqi@0: default: aoqi@0: if (cmpxy == NULL) aoqi@0: cmpxy = ideal_cmpxy; aoqi@0: best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt)); aoqi@0: // and fall through: aoqi@0: case BoolTest::lt: // x < y aoqi@0: case BoolTest::le: // x <= y aoqi@0: answer_if_true = (want_max ? yvalue : xvalue); aoqi@0: answer_if_false = (want_max ? xvalue : yvalue); aoqi@0: break; aoqi@0: case BoolTest::gt: // x > y aoqi@0: case BoolTest::ge: // x >= y aoqi@0: answer_if_true = (want_max ? xvalue : yvalue); aoqi@0: answer_if_false = (want_max ? yvalue : xvalue); aoqi@0: break; aoqi@0: } aoqi@0: aoqi@0: jint hi, lo; aoqi@0: if (want_max) { aoqi@0: // We can sharpen the minimum. aoqi@0: hi = MAX2(txvalue->_hi, tyvalue->_hi); aoqi@0: lo = MAX2(txvalue->_lo, tyvalue->_lo); aoqi@0: } else { aoqi@0: // We can sharpen the maximum. aoqi@0: hi = MIN2(txvalue->_hi, tyvalue->_hi); aoqi@0: lo = MIN2(txvalue->_lo, tyvalue->_lo); aoqi@0: } aoqi@0: aoqi@0: // Use a flow-free graph structure, to avoid creating excess control edges aoqi@0: // which could hinder other optimizations. aoqi@0: // Since Math.min/max is often used with arraycopy, we want aoqi@0: // tightly_coupled_allocation to be able to see beyond min/max expressions. aoqi@0: Node* cmov = CMoveNode::make(C, NULL, best_bol, aoqi@0: answer_if_false, answer_if_true, aoqi@0: TypeInt::make(lo, hi, widen)); aoqi@0: aoqi@0: return _gvn.transform(cmov); aoqi@0: aoqi@0: /* aoqi@0: // This is not as desirable as it may seem, since Min and Max aoqi@0: // nodes do not have a full set of optimizations. aoqi@0: // And they would interfere, anyway, with 'if' optimizations aoqi@0: // and with CMoveI canonical forms. aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_min: aoqi@0: result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; aoqi@0: case vmIntrinsics::_max: aoqi@0: result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; aoqi@0: default: aoqi@0: ShouldNotReachHere(); aoqi@0: } aoqi@0: */ aoqi@0: } aoqi@0: aoqi@0: inline int aoqi@0: LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { aoqi@0: const TypePtr* base_type = TypePtr::NULL_PTR; aoqi@0: if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); aoqi@0: if (base_type == NULL) { aoqi@0: // Unknown type. aoqi@0: return Type::AnyPtr; aoqi@0: } else if (base_type == TypePtr::NULL_PTR) { aoqi@0: // Since this is a NULL+long form, we have to switch to a rawptr. aoqi@0: base = _gvn.transform(new (C) CastX2PNode(offset)); aoqi@0: offset = MakeConX(0); aoqi@0: return Type::RawPtr; aoqi@0: } else if (base_type->base() == Type::RawPtr) { aoqi@0: return Type::RawPtr; aoqi@0: } else if (base_type->isa_oopptr()) { aoqi@0: // Base is never null => always a heap address. aoqi@0: if (base_type->ptr() == TypePtr::NotNull) { aoqi@0: return Type::OopPtr; aoqi@0: } aoqi@0: // Offset is small => always a heap address. aoqi@0: const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); aoqi@0: if (offset_type != NULL && aoqi@0: base_type->offset() == 0 && // (should always be?) aoqi@0: offset_type->_lo >= 0 && aoqi@0: !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { aoqi@0: return Type::OopPtr; aoqi@0: } aoqi@0: // Otherwise, it might either be oop+off or NULL+addr. aoqi@0: return Type::AnyPtr; aoqi@0: } else { aoqi@0: // No information: aoqi@0: return Type::AnyPtr; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { aoqi@0: int kind = classify_unsafe_addr(base, offset); aoqi@0: if (kind == Type::RawPtr) { aoqi@0: return basic_plus_adr(top(), base, offset); aoqi@0: } else { aoqi@0: return basic_plus_adr(base, offset); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: //--------------------------inline_number_methods----------------------------- aoqi@0: // inline int Integer.numberOfLeadingZeros(int) aoqi@0: // inline int Long.numberOfLeadingZeros(long) aoqi@0: // aoqi@0: // inline int Integer.numberOfTrailingZeros(int) aoqi@0: // inline int Long.numberOfTrailingZeros(long) aoqi@0: // aoqi@0: // inline int Integer.bitCount(int) aoqi@0: // inline int Long.bitCount(long) aoqi@0: // aoqi@0: // inline char Character.reverseBytes(char) aoqi@0: // inline short Short.reverseBytes(short) aoqi@0: // inline int Integer.reverseBytes(int) aoqi@0: // inline long Long.reverseBytes(long) aoqi@0: bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { aoqi@0: Node* arg = argument(0); aoqi@0: Node* n; aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break; aoqi@0: case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break; aoqi@0: case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break; aoqi@0: case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break; aoqi@0: case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break; aoqi@0: case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break; aoqi@0: case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break; aoqi@0: case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break; aoqi@0: case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break; aoqi@0: case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break; aoqi@0: default: fatal_unexpected_iid(id); break; aoqi@0: } aoqi@0: set_result(_gvn.transform(n)); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //----------------------------inline_unsafe_access---------------------------- aoqi@0: aoqi@0: const static BasicType T_ADDRESS_HOLDER = T_LONG; aoqi@0: aoqi@0: // Helper that guards and inserts a pre-barrier. aoqi@0: void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, aoqi@0: Node* pre_val, bool need_mem_bar) { aoqi@0: // We could be accessing the referent field of a reference object. If so, when G1 aoqi@0: // is enabled, we need to log the value in the referent field in an SATB buffer. aoqi@0: // This routine performs some compile time filters and generates suitable aoqi@0: // runtime filters that guard the pre-barrier code. aoqi@0: // Also add memory barrier for non volatile load from the referent field aoqi@0: // to prevent commoning of loads across safepoint. aoqi@0: if (!UseG1GC && !need_mem_bar) aoqi@0: return; aoqi@0: aoqi@0: // Some compile time checks. aoqi@0: aoqi@0: // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? aoqi@0: const TypeX* otype = offset->find_intptr_t_type(); aoqi@0: if (otype != NULL && otype->is_con() && aoqi@0: otype->get_con() != java_lang_ref_Reference::referent_offset) { aoqi@0: // Constant offset but not the reference_offset so just return aoqi@0: return; aoqi@0: } aoqi@0: aoqi@0: // We only need to generate the runtime guards for instances. aoqi@0: const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); aoqi@0: if (btype != NULL) { aoqi@0: if (btype->isa_aryptr()) { aoqi@0: // Array type so nothing to do aoqi@0: return; aoqi@0: } aoqi@0: aoqi@0: const TypeInstPtr* itype = btype->isa_instptr(); aoqi@0: if (itype != NULL) { aoqi@0: // Can the klass of base_oop be statically determined to be aoqi@0: // _not_ a sub-class of Reference and _not_ Object? aoqi@0: ciKlass* klass = itype->klass(); aoqi@0: if ( klass->is_loaded() && aoqi@0: !klass->is_subtype_of(env()->Reference_klass()) && aoqi@0: !env()->Object_klass()->is_subtype_of(klass)) { aoqi@0: return; aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // The compile time filters did not reject base_oop/offset so aoqi@0: // we need to generate the following runtime filters aoqi@0: // aoqi@0: // if (offset == java_lang_ref_Reference::_reference_offset) { aoqi@0: // if (instance_of(base, java.lang.ref.Reference)) { aoqi@0: // pre_barrier(_, pre_val, ...); aoqi@0: // } aoqi@0: // } aoqi@0: aoqi@0: float likely = PROB_LIKELY( 0.999); aoqi@0: float unlikely = PROB_UNLIKELY(0.999); aoqi@0: aoqi@0: IdealKit ideal(this); aoqi@0: #define __ ideal. aoqi@0: aoqi@0: Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); aoqi@0: aoqi@0: __ if_then(offset, BoolTest::eq, referent_off, unlikely); { aoqi@0: // Update graphKit memory and control from IdealKit. aoqi@0: sync_kit(ideal); aoqi@0: aoqi@0: Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); aoqi@0: Node* is_instof = gen_instanceof(base_oop, ref_klass_con); aoqi@0: aoqi@0: // Update IdealKit memory and control from graphKit. aoqi@0: __ sync_kit(this); aoqi@0: aoqi@0: Node* one = __ ConI(1); aoqi@0: // is_instof == 0 if base_oop == NULL aoqi@0: __ if_then(is_instof, BoolTest::eq, one, unlikely); { aoqi@0: aoqi@0: // Update graphKit from IdeakKit. aoqi@0: sync_kit(ideal); aoqi@0: aoqi@0: // Use the pre-barrier to record the value in the referent field aoqi@0: pre_barrier(false /* do_load */, aoqi@0: __ ctrl(), aoqi@0: NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, aoqi@0: pre_val /* pre_val */, aoqi@0: T_OBJECT); aoqi@0: if (need_mem_bar) { aoqi@0: // Add memory barrier to prevent commoning reads from this field aoqi@0: // across safepoint since GC can change its value. aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: } aoqi@0: // Update IdealKit from graphKit. aoqi@0: __ sync_kit(this); aoqi@0: aoqi@0: } __ end_if(); // _ref_type != ref_none aoqi@0: } __ end_if(); // offset == referent_offset aoqi@0: aoqi@0: // Final sync IdealKit and GraphKit. aoqi@0: final_sync(ideal); aoqi@0: #undef __ aoqi@0: } aoqi@0: aoqi@0: aoqi@0: // Interpret Unsafe.fieldOffset cookies correctly: aoqi@0: extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); aoqi@0: aoqi@0: const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { aoqi@0: // Attempt to infer a sharper value type from the offset and base type. aoqi@0: ciKlass* sharpened_klass = NULL; aoqi@0: aoqi@0: // See if it is an instance field, with an object type. aoqi@0: if (alias_type->field() != NULL) { aoqi@0: assert(!is_native_ptr, "native pointer op cannot use a java address"); aoqi@0: if (alias_type->field()->type()->is_klass()) { aoqi@0: sharpened_klass = alias_type->field()->type()->as_klass(); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // See if it is a narrow oop array. aoqi@0: if (adr_type->isa_aryptr()) { aoqi@0: if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { aoqi@0: const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); aoqi@0: if (elem_type != NULL) { aoqi@0: sharpened_klass = elem_type->klass(); aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // The sharpened class might be unloaded if there is no class loader aoqi@0: // contraint in place. aoqi@0: if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { aoqi@0: const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); aoqi@0: aoqi@0: #ifndef PRODUCT aoqi@0: if (C->print_intrinsics() || C->print_inlining()) { aoqi@0: tty->print(" from base type: "); adr_type->dump(); aoqi@0: tty->print(" sharpened value: "); tjp->dump(); aoqi@0: } aoqi@0: #endif aoqi@0: // Sharpen the value type. aoqi@0: return tjp; aoqi@0: } aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { aoqi@0: if (callee()->is_static()) return false; // caller must have the capability! aoqi@0: aoqi@0: #ifndef PRODUCT aoqi@0: { aoqi@0: ResourceMark rm; aoqi@0: // Check the signatures. aoqi@0: ciSignature* sig = callee()->signature(); aoqi@0: #ifdef ASSERT aoqi@0: if (!is_store) { aoqi@0: // Object getObject(Object base, int/long offset), etc. aoqi@0: BasicType rtype = sig->return_type()->basic_type(); aoqi@0: if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) aoqi@0: rtype = T_ADDRESS; // it is really a C void* aoqi@0: assert(rtype == type, "getter must return the expected value"); aoqi@0: if (!is_native_ptr) { aoqi@0: assert(sig->count() == 2, "oop getter has 2 arguments"); aoqi@0: assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); aoqi@0: assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); aoqi@0: } else { aoqi@0: assert(sig->count() == 1, "native getter has 1 argument"); aoqi@0: assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); aoqi@0: } aoqi@0: } else { aoqi@0: // void putObject(Object base, int/long offset, Object x), etc. aoqi@0: assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); aoqi@0: if (!is_native_ptr) { aoqi@0: assert(sig->count() == 3, "oop putter has 3 arguments"); aoqi@0: assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); aoqi@0: assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); aoqi@0: } else { aoqi@0: assert(sig->count() == 2, "native putter has 2 arguments"); aoqi@0: assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); aoqi@0: } aoqi@0: BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); aoqi@0: if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) aoqi@0: vtype = T_ADDRESS; // it is really a C void* aoqi@0: assert(vtype == type, "putter must accept the expected value"); aoqi@0: } aoqi@0: #endif // ASSERT aoqi@0: } aoqi@0: #endif //PRODUCT aoqi@0: aoqi@0: C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". aoqi@0: aoqi@0: Node* receiver = argument(0); // type: oop aoqi@0: aoqi@0: // Build address expression. See the code in inline_unsafe_prefetch. aoqi@0: Node* adr; aoqi@0: Node* heap_base_oop = top(); aoqi@0: Node* offset = top(); aoqi@0: Node* val; aoqi@0: aoqi@0: if (!is_native_ptr) { aoqi@0: // The base is either a Java object or a value produced by Unsafe.staticFieldBase aoqi@0: Node* base = argument(1); // type: oop aoqi@0: // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset aoqi@0: offset = argument(2); // type: long aoqi@0: // We currently rely on the cookies produced by Unsafe.xxxFieldOffset aoqi@0: // to be plain byte offsets, which are also the same as those accepted aoqi@0: // by oopDesc::field_base. aoqi@0: assert(Unsafe_field_offset_to_byte_offset(11) == 11, aoqi@0: "fieldOffset must be byte-scaled"); aoqi@0: // 32-bit machines ignore the high half! aoqi@0: offset = ConvL2X(offset); aoqi@0: adr = make_unsafe_address(base, offset); aoqi@0: heap_base_oop = base; aoqi@0: val = is_store ? argument(4) : NULL; aoqi@0: } else { aoqi@0: Node* ptr = argument(1); // type: long aoqi@0: ptr = ConvL2X(ptr); // adjust Java long to machine word aoqi@0: adr = make_unsafe_address(NULL, ptr); aoqi@0: val = is_store ? argument(3) : NULL; aoqi@0: } aoqi@0: aoqi@0: const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); aoqi@0: aoqi@0: // First guess at the value type. aoqi@0: const Type *value_type = Type::get_const_basic_type(type); aoqi@0: aoqi@0: // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, aoqi@0: // there was not enough information to nail it down. aoqi@0: Compile::AliasType* alias_type = C->alias_type(adr_type); aoqi@0: assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); aoqi@0: aoqi@0: // We will need memory barriers unless we can determine a unique aoqi@0: // alias category for this reference. (Note: If for some reason aoqi@0: // the barriers get omitted and the unsafe reference begins to "pollute" aoqi@0: // the alias analysis of the rest of the graph, either Compile::can_alias aoqi@0: // or Compile::must_alias will throw a diagnostic assert.) aoqi@0: bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); aoqi@0: aoqi@0: // If we are reading the value of the referent field of a Reference aoqi@0: // object (either by using Unsafe directly or through reflection) aoqi@0: // then, if G1 is enabled, we need to record the referent in an aoqi@0: // SATB log buffer using the pre-barrier mechanism. aoqi@0: // Also we need to add memory barrier to prevent commoning reads aoqi@0: // from this field across safepoint since GC can change its value. aoqi@0: bool need_read_barrier = !is_native_ptr && !is_store && aoqi@0: offset != top() && heap_base_oop != top(); aoqi@0: aoqi@0: if (!is_store && type == T_OBJECT) { aoqi@0: const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); aoqi@0: if (tjp != NULL) { aoqi@0: value_type = tjp; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: receiver = null_check(receiver); aoqi@0: if (stopped()) { aoqi@0: return true; aoqi@0: } aoqi@0: // Heap pointers get a null-check from the interpreter, aoqi@0: // as a courtesy. However, this is not guaranteed by Unsafe, aoqi@0: // and it is not possible to fully distinguish unintended nulls aoqi@0: // from intended ones in this API. aoqi@0: aoqi@0: if (is_volatile) { aoqi@0: // We need to emit leading and trailing CPU membars (see below) in aoqi@0: // addition to memory membars when is_volatile. This is a little aoqi@0: // too strong, but avoids the need to insert per-alias-type aoqi@0: // volatile membars (for stores; compare Parse::do_put_xxx), which aoqi@0: // we cannot do effectively here because we probably only have a aoqi@0: // rough approximation of type. aoqi@0: need_mem_bar = true; aoqi@0: // For Stores, place a memory ordering barrier now. aoqi@0: if (is_store) { aoqi@0: insert_mem_bar(Op_MemBarRelease); aoqi@0: } else { aoqi@0: if (support_IRIW_for_not_multiple_copy_atomic_cpu) { aoqi@0: insert_mem_bar(Op_MemBarVolatile); aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // Memory barrier to prevent normal and 'unsafe' accesses from aoqi@0: // bypassing each other. Happens after null checks, so the aoqi@0: // exception paths do not take memory state from the memory barrier, aoqi@0: // so there's no problems making a strong assert about mixing users aoqi@0: // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar aoqi@0: // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. aoqi@0: if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: aoqi@0: if (!is_store) { aoqi@0: Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile); aoqi@0: // load value aoqi@0: switch (type) { aoqi@0: case T_BOOLEAN: aoqi@0: case T_CHAR: aoqi@0: case T_BYTE: aoqi@0: case T_SHORT: aoqi@0: case T_INT: aoqi@0: case T_LONG: aoqi@0: case T_FLOAT: aoqi@0: case T_DOUBLE: aoqi@0: break; aoqi@0: case T_OBJECT: aoqi@0: if (need_read_barrier) { aoqi@0: insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); aoqi@0: } aoqi@0: break; aoqi@0: case T_ADDRESS: aoqi@0: // Cast to an int type. aoqi@0: p = _gvn.transform(new (C) CastP2XNode(NULL, p)); aoqi@0: p = ConvX2UL(p); aoqi@0: break; aoqi@0: default: aoqi@0: fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); aoqi@0: break; aoqi@0: } aoqi@0: // The load node has the control of the preceding MemBarCPUOrder. All aoqi@0: // following nodes will have the control of the MemBarCPUOrder inserted at aoqi@0: // the end of this method. So, pushing the load onto the stack at a later aoqi@0: // point is fine. aoqi@0: set_result(p); aoqi@0: } else { aoqi@0: // place effect of store into memory aoqi@0: switch (type) { aoqi@0: case T_DOUBLE: aoqi@0: val = dstore_rounding(val); aoqi@0: break; aoqi@0: case T_ADDRESS: aoqi@0: // Repackage the long as a pointer. aoqi@0: val = ConvL2X(val); aoqi@0: val = _gvn.transform(new (C) CastX2PNode(val)); aoqi@0: break; aoqi@0: } aoqi@0: aoqi@0: MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; aoqi@0: if (type != T_OBJECT ) { aoqi@0: (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); aoqi@0: } else { aoqi@0: // Possibly an oop being stored to Java heap or native memory aoqi@0: if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { aoqi@0: // oop to Java heap. aoqi@0: (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); aoqi@0: } else { aoqi@0: // We can't tell at compile time if we are storing in the Java heap or outside aoqi@0: // of it. So we need to emit code to conditionally do the proper type of aoqi@0: // store. aoqi@0: aoqi@0: IdealKit ideal(this); aoqi@0: #define __ ideal. aoqi@0: // QQQ who knows what probability is here?? aoqi@0: __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { aoqi@0: // Sync IdealKit and graphKit. aoqi@0: sync_kit(ideal); aoqi@0: Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); aoqi@0: // Update IdealKit memory. aoqi@0: __ sync_kit(this); aoqi@0: } __ else_(); { aoqi@0: __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); aoqi@0: } __ end_if(); aoqi@0: // Final sync IdealKit and GraphKit. aoqi@0: final_sync(ideal); aoqi@0: #undef __ aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: if (is_volatile) { aoqi@0: if (!is_store) { aoqi@0: insert_mem_bar(Op_MemBarAcquire); aoqi@0: } else { aoqi@0: if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { aoqi@0: insert_mem_bar(Op_MemBarVolatile); aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //----------------------------inline_unsafe_prefetch---------------------------- aoqi@0: aoqi@0: bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { aoqi@0: #ifndef PRODUCT aoqi@0: { aoqi@0: ResourceMark rm; aoqi@0: // Check the signatures. aoqi@0: ciSignature* sig = callee()->signature(); aoqi@0: #ifdef ASSERT aoqi@0: // Object getObject(Object base, int/long offset), etc. aoqi@0: BasicType rtype = sig->return_type()->basic_type(); aoqi@0: if (!is_native_ptr) { aoqi@0: assert(sig->count() == 2, "oop prefetch has 2 arguments"); aoqi@0: assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); aoqi@0: assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); aoqi@0: } else { aoqi@0: assert(sig->count() == 1, "native prefetch has 1 argument"); aoqi@0: assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); aoqi@0: } aoqi@0: #endif // ASSERT aoqi@0: } aoqi@0: #endif // !PRODUCT aoqi@0: aoqi@0: C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". aoqi@0: aoqi@0: const int idx = is_static ? 0 : 1; aoqi@0: if (!is_static) { aoqi@0: null_check_receiver(); aoqi@0: if (stopped()) { aoqi@0: return true; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // Build address expression. See the code in inline_unsafe_access. aoqi@0: Node *adr; aoqi@0: if (!is_native_ptr) { aoqi@0: // The base is either a Java object or a value produced by Unsafe.staticFieldBase aoqi@0: Node* base = argument(idx + 0); // type: oop aoqi@0: // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset aoqi@0: Node* offset = argument(idx + 1); // type: long aoqi@0: // We currently rely on the cookies produced by Unsafe.xxxFieldOffset aoqi@0: // to be plain byte offsets, which are also the same as those accepted aoqi@0: // by oopDesc::field_base. aoqi@0: assert(Unsafe_field_offset_to_byte_offset(11) == 11, aoqi@0: "fieldOffset must be byte-scaled"); aoqi@0: // 32-bit machines ignore the high half! aoqi@0: offset = ConvL2X(offset); aoqi@0: adr = make_unsafe_address(base, offset); aoqi@0: } else { aoqi@0: Node* ptr = argument(idx + 0); // type: long aoqi@0: ptr = ConvL2X(ptr); // adjust Java long to machine word aoqi@0: adr = make_unsafe_address(NULL, ptr); aoqi@0: } aoqi@0: aoqi@0: // Generate the read or write prefetch aoqi@0: Node *prefetch; aoqi@0: if (is_store) { aoqi@0: prefetch = new (C) PrefetchWriteNode(i_o(), adr); aoqi@0: } else { aoqi@0: prefetch = new (C) PrefetchReadNode(i_o(), adr); aoqi@0: } aoqi@0: prefetch->init_req(0, control()); aoqi@0: set_i_o(_gvn.transform(prefetch)); aoqi@0: aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //----------------------------inline_unsafe_load_store---------------------------- aoqi@0: // This method serves a couple of different customers (depending on LoadStoreKind): aoqi@0: // aoqi@0: // LS_cmpxchg: aoqi@0: // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); aoqi@0: // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); aoqi@0: // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); aoqi@0: // aoqi@0: // LS_xadd: aoqi@0: // public int getAndAddInt( Object o, long offset, int delta) aoqi@0: // public long getAndAddLong(Object o, long offset, long delta) aoqi@0: // aoqi@0: // LS_xchg: aoqi@0: // int getAndSet(Object o, long offset, int newValue) aoqi@0: // long getAndSet(Object o, long offset, long newValue) aoqi@0: // Object getAndSet(Object o, long offset, Object newValue) aoqi@0: // aoqi@0: bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { aoqi@0: // This basic scheme here is the same as inline_unsafe_access, but aoqi@0: // differs in enough details that combining them would make the code aoqi@0: // overly confusing. (This is a true fact! I originally combined aoqi@0: // them, but even I was confused by it!) As much code/comments as aoqi@0: // possible are retained from inline_unsafe_access though to make aoqi@0: // the correspondences clearer. - dl aoqi@0: aoqi@0: if (callee()->is_static()) return false; // caller must have the capability! aoqi@0: aoqi@0: #ifndef PRODUCT aoqi@0: BasicType rtype; aoqi@0: { aoqi@0: ResourceMark rm; aoqi@0: // Check the signatures. aoqi@0: ciSignature* sig = callee()->signature(); aoqi@0: rtype = sig->return_type()->basic_type(); aoqi@0: if (kind == LS_xadd || kind == LS_xchg) { aoqi@0: // Check the signatures. aoqi@0: #ifdef ASSERT aoqi@0: assert(rtype == type, "get and set must return the expected type"); aoqi@0: assert(sig->count() == 3, "get and set has 3 arguments"); aoqi@0: assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); aoqi@0: assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); aoqi@0: assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); aoqi@0: #endif // ASSERT aoqi@0: } else if (kind == LS_cmpxchg) { aoqi@0: // Check the signatures. aoqi@0: #ifdef ASSERT aoqi@0: assert(rtype == T_BOOLEAN, "CAS must return boolean"); aoqi@0: assert(sig->count() == 4, "CAS has 4 arguments"); aoqi@0: assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); aoqi@0: assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); aoqi@0: #endif // ASSERT aoqi@0: } else { aoqi@0: ShouldNotReachHere(); aoqi@0: } aoqi@0: } aoqi@0: #endif //PRODUCT aoqi@0: aoqi@0: C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". aoqi@0: aoqi@0: // Get arguments: aoqi@0: Node* receiver = NULL; aoqi@0: Node* base = NULL; aoqi@0: Node* offset = NULL; aoqi@0: Node* oldval = NULL; aoqi@0: Node* newval = NULL; aoqi@0: if (kind == LS_cmpxchg) { aoqi@0: const bool two_slot_type = type2size[type] == 2; aoqi@0: receiver = argument(0); // type: oop aoqi@0: base = argument(1); // type: oop aoqi@0: offset = argument(2); // type: long aoqi@0: oldval = argument(4); // type: oop, int, or long aoqi@0: newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long aoqi@0: } else if (kind == LS_xadd || kind == LS_xchg){ aoqi@0: receiver = argument(0); // type: oop aoqi@0: base = argument(1); // type: oop aoqi@0: offset = argument(2); // type: long aoqi@0: oldval = NULL; aoqi@0: newval = argument(4); // type: oop, int, or long aoqi@0: } aoqi@0: aoqi@0: // Null check receiver. aoqi@0: receiver = null_check(receiver); aoqi@0: if (stopped()) { aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: // Build field offset expression. aoqi@0: // We currently rely on the cookies produced by Unsafe.xxxFieldOffset aoqi@0: // to be plain byte offsets, which are also the same as those accepted aoqi@0: // by oopDesc::field_base. aoqi@0: assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); aoqi@0: // 32-bit machines ignore the high half of long offsets aoqi@0: offset = ConvL2X(offset); aoqi@0: Node* adr = make_unsafe_address(base, offset); aoqi@0: const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); aoqi@0: aoqi@0: // For CAS, unlike inline_unsafe_access, there seems no point in aoqi@0: // trying to refine types. Just use the coarse types here. aoqi@0: const Type *value_type = Type::get_const_basic_type(type); aoqi@0: Compile::AliasType* alias_type = C->alias_type(adr_type); aoqi@0: assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); aoqi@0: aoqi@0: if (kind == LS_xchg && type == T_OBJECT) { aoqi@0: const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); aoqi@0: if (tjp != NULL) { aoqi@0: value_type = tjp; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: int alias_idx = C->get_alias_index(adr_type); aoqi@0: aoqi@0: // Memory-model-wise, a LoadStore acts like a little synchronized aoqi@0: // block, so needs barriers on each side. These don't translate aoqi@0: // into actual barriers on most machines, but we still need rest of aoqi@0: // compiler to respect ordering. aoqi@0: aoqi@0: insert_mem_bar(Op_MemBarRelease); aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: aoqi@0: // 4984716: MemBars must be inserted before this aoqi@0: // memory node in order to avoid a false aoqi@0: // dependency which will confuse the scheduler. aoqi@0: Node *mem = memory(alias_idx); aoqi@0: aoqi@0: // For now, we handle only those cases that actually exist: ints, aoqi@0: // longs, and Object. Adding others should be straightforward. aoqi@0: Node* load_store; aoqi@0: switch(type) { aoqi@0: case T_INT: aoqi@0: if (kind == LS_xadd) { aoqi@0: load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type)); aoqi@0: } else if (kind == LS_xchg) { aoqi@0: load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type)); aoqi@0: } else if (kind == LS_cmpxchg) { aoqi@0: load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval)); aoqi@0: } else { aoqi@0: ShouldNotReachHere(); aoqi@0: } aoqi@0: break; aoqi@0: case T_LONG: aoqi@0: if (kind == LS_xadd) { aoqi@0: load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type)); aoqi@0: } else if (kind == LS_xchg) { aoqi@0: load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type)); aoqi@0: } else if (kind == LS_cmpxchg) { aoqi@0: load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); aoqi@0: } else { aoqi@0: ShouldNotReachHere(); aoqi@0: } aoqi@0: break; aoqi@0: case T_OBJECT: aoqi@0: // Transformation of a value which could be NULL pointer (CastPP #NULL) aoqi@0: // could be delayed during Parse (for example, in adjust_map_after_if()). aoqi@0: // Execute transformation here to avoid barrier generation in such case. aoqi@0: if (_gvn.type(newval) == TypePtr::NULL_PTR) aoqi@0: newval = _gvn.makecon(TypePtr::NULL_PTR); aoqi@0: aoqi@0: // Reference stores need a store barrier. aoqi@0: if (kind == LS_xchg) { aoqi@0: // If pre-barrier must execute before the oop store, old value will require do_load here. aoqi@0: if (!can_move_pre_barrier()) { aoqi@0: pre_barrier(true /* do_load*/, aoqi@0: control(), base, adr, alias_idx, newval, value_type->make_oopptr(), aoqi@0: NULL /* pre_val*/, aoqi@0: T_OBJECT); aoqi@0: } // Else move pre_barrier to use load_store value, see below. aoqi@0: } else if (kind == LS_cmpxchg) { aoqi@0: // Same as for newval above: aoqi@0: if (_gvn.type(oldval) == TypePtr::NULL_PTR) { aoqi@0: oldval = _gvn.makecon(TypePtr::NULL_PTR); aoqi@0: } aoqi@0: // The only known value which might get overwritten is oldval. aoqi@0: pre_barrier(false /* do_load */, aoqi@0: control(), NULL, NULL, max_juint, NULL, NULL, aoqi@0: oldval /* pre_val */, aoqi@0: T_OBJECT); aoqi@0: } else { aoqi@0: ShouldNotReachHere(); aoqi@0: } aoqi@0: aoqi@0: #ifdef _LP64 aoqi@0: if (adr->bottom_type()->is_ptr_to_narrowoop()) { aoqi@0: Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); aoqi@0: if (kind == LS_xchg) { aoqi@0: load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr, aoqi@0: newval_enc, adr_type, value_type->make_narrowoop())); aoqi@0: } else { aoqi@0: assert(kind == LS_cmpxchg, "wrong LoadStore operation"); aoqi@0: Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); aoqi@0: load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr, aoqi@0: newval_enc, oldval_enc)); aoqi@0: } aoqi@0: } else aoqi@0: #endif aoqi@0: { aoqi@0: if (kind == LS_xchg) { aoqi@0: load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); aoqi@0: } else { aoqi@0: assert(kind == LS_cmpxchg, "wrong LoadStore operation"); aoqi@0: load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); aoqi@0: } aoqi@0: } aoqi@0: post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); aoqi@0: break; aoqi@0: default: aoqi@0: fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); aoqi@0: break; aoqi@0: } aoqi@0: aoqi@0: // SCMemProjNodes represent the memory state of a LoadStore. Their aoqi@0: // main role is to prevent LoadStore nodes from being optimized away aoqi@0: // when their results aren't used. aoqi@0: Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store)); aoqi@0: set_memory(proj, alias_idx); aoqi@0: aoqi@0: if (type == T_OBJECT && kind == LS_xchg) { aoqi@0: #ifdef _LP64 aoqi@0: if (adr->bottom_type()->is_ptr_to_narrowoop()) { aoqi@0: load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type())); aoqi@0: } aoqi@0: #endif aoqi@0: if (can_move_pre_barrier()) { aoqi@0: // Don't need to load pre_val. The old value is returned by load_store. aoqi@0: // The pre_barrier can execute after the xchg as long as no safepoint aoqi@0: // gets inserted between them. aoqi@0: pre_barrier(false /* do_load */, aoqi@0: control(), NULL, NULL, max_juint, NULL, NULL, aoqi@0: load_store /* pre_val */, aoqi@0: T_OBJECT); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // Add the trailing membar surrounding the access aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: insert_mem_bar(Op_MemBarAcquire); aoqi@0: aoqi@0: assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); aoqi@0: set_result(load_store); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //----------------------------inline_unsafe_ordered_store---------------------- aoqi@0: // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); aoqi@0: // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); aoqi@0: // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); aoqi@0: bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { aoqi@0: // This is another variant of inline_unsafe_access, differing in aoqi@0: // that it always issues store-store ("release") barrier and ensures aoqi@0: // store-atomicity (which only matters for "long"). aoqi@0: aoqi@0: if (callee()->is_static()) return false; // caller must have the capability! aoqi@0: aoqi@0: #ifndef PRODUCT aoqi@0: { aoqi@0: ResourceMark rm; aoqi@0: // Check the signatures. aoqi@0: ciSignature* sig = callee()->signature(); aoqi@0: #ifdef ASSERT aoqi@0: BasicType rtype = sig->return_type()->basic_type(); aoqi@0: assert(rtype == T_VOID, "must return void"); aoqi@0: assert(sig->count() == 3, "has 3 arguments"); aoqi@0: assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); aoqi@0: assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); aoqi@0: #endif // ASSERT aoqi@0: } aoqi@0: #endif //PRODUCT aoqi@0: aoqi@0: C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". aoqi@0: aoqi@0: // Get arguments: aoqi@0: Node* receiver = argument(0); // type: oop aoqi@0: Node* base = argument(1); // type: oop aoqi@0: Node* offset = argument(2); // type: long aoqi@0: Node* val = argument(4); // type: oop, int, or long aoqi@0: aoqi@0: // Null check receiver. aoqi@0: receiver = null_check(receiver); aoqi@0: if (stopped()) { aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: // Build field offset expression. aoqi@0: assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); aoqi@0: // 32-bit machines ignore the high half of long offsets aoqi@0: offset = ConvL2X(offset); aoqi@0: Node* adr = make_unsafe_address(base, offset); aoqi@0: const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); aoqi@0: const Type *value_type = Type::get_const_basic_type(type); aoqi@0: Compile::AliasType* alias_type = C->alias_type(adr_type); aoqi@0: aoqi@0: insert_mem_bar(Op_MemBarRelease); aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: // Ensure that the store is atomic for longs: aoqi@0: const bool require_atomic_access = true; aoqi@0: Node* store; aoqi@0: if (type == T_OBJECT) // reference stores need a store barrier. aoqi@0: store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); aoqi@0: else { aoqi@0: store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); aoqi@0: } aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { aoqi@0: // Regardless of form, don't allow previous ld/st to move down, aoqi@0: // then issue acquire, release, or volatile mem_bar. aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: switch(id) { aoqi@0: case vmIntrinsics::_loadFence: aoqi@0: insert_mem_bar(Op_LoadFence); aoqi@0: return true; aoqi@0: case vmIntrinsics::_storeFence: aoqi@0: insert_mem_bar(Op_StoreFence); aoqi@0: return true; aoqi@0: case vmIntrinsics::_fullFence: aoqi@0: insert_mem_bar(Op_MemBarVolatile); aoqi@0: return true; aoqi@0: default: aoqi@0: fatal_unexpected_iid(id); aoqi@0: return false; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::klass_needs_init_guard(Node* kls) { aoqi@0: if (!kls->is_Con()) { aoqi@0: return true; aoqi@0: } aoqi@0: const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); aoqi@0: if (klsptr == NULL) { aoqi@0: return true; aoqi@0: } aoqi@0: ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); aoqi@0: // don't need a guard for a klass that is already initialized aoqi@0: return !ik->is_initialized(); aoqi@0: } aoqi@0: aoqi@0: //----------------------------inline_unsafe_allocate--------------------------- aoqi@0: // public native Object sun.misc.Unsafe.allocateInstance(Class cls); aoqi@0: bool LibraryCallKit::inline_unsafe_allocate() { aoqi@0: if (callee()->is_static()) return false; // caller must have the capability! aoqi@0: aoqi@0: null_check_receiver(); // null-check, then ignore aoqi@0: Node* cls = null_check(argument(1)); aoqi@0: if (stopped()) return true; aoqi@0: aoqi@0: Node* kls = load_klass_from_mirror(cls, false, NULL, 0); aoqi@0: kls = null_check(kls); aoqi@0: if (stopped()) return true; // argument was like int.class aoqi@0: aoqi@0: Node* test = NULL; aoqi@0: if (LibraryCallKit::klass_needs_init_guard(kls)) { aoqi@0: // Note: The argument might still be an illegal value like aoqi@0: // Serializable.class or Object[].class. The runtime will handle it. aoqi@0: // But we must make an explicit check for initialization. aoqi@0: Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); aoqi@0: // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler aoqi@0: // can generate code to load it as unsigned byte. aoqi@0: Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); aoqi@0: Node* bits = intcon(InstanceKlass::fully_initialized); aoqi@0: test = _gvn.transform(new (C) SubINode(inst, bits)); aoqi@0: // The 'test' is non-zero if we need to take a slow path. aoqi@0: } aoqi@0: aoqi@0: Node* obj = new_instance(kls, test); aoqi@0: set_result(obj); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: #ifdef TRACE_HAVE_INTRINSICS aoqi@0: /* aoqi@0: * oop -> myklass aoqi@0: * myklass->trace_id |= USED aoqi@0: * return myklass->trace_id & ~0x3 aoqi@0: */ aoqi@0: bool LibraryCallKit::inline_native_classID() { aoqi@0: null_check_receiver(); // null-check, then ignore aoqi@0: Node* cls = null_check(argument(1), T_OBJECT); aoqi@0: Node* kls = load_klass_from_mirror(cls, false, NULL, 0); aoqi@0: kls = null_check(kls, T_OBJECT); aoqi@0: ByteSize offset = TRACE_ID_OFFSET; aoqi@0: Node* insp = basic_plus_adr(kls, in_bytes(offset)); aoqi@0: Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); aoqi@0: Node* bits = longcon(~0x03l); // ignore bit 0 & 1 aoqi@0: Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits)); aoqi@0: Node* clsused = longcon(0x01l); // set the class bit aoqi@0: Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused)); aoqi@0: aoqi@0: const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); aoqi@0: store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); aoqi@0: set_result(andl); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_native_threadID() { aoqi@0: Node* tls_ptr = NULL; aoqi@0: Node* cur_thr = generate_current_thread(tls_ptr); aoqi@0: Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); aoqi@0: Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); aoqi@0: p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); aoqi@0: aoqi@0: Node* threadid = NULL; aoqi@0: size_t thread_id_size = OSThread::thread_id_size(); aoqi@0: if (thread_id_size == (size_t) BytesPerLong) { aoqi@0: threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); aoqi@0: } else if (thread_id_size == (size_t) BytesPerInt) { aoqi@0: threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); aoqi@0: } else { aoqi@0: ShouldNotReachHere(); aoqi@0: } aoqi@0: set_result(threadid); aoqi@0: return true; aoqi@0: } aoqi@0: #endif aoqi@0: aoqi@0: //------------------------inline_native_time_funcs-------------- aoqi@0: // inline code for System.currentTimeMillis() and System.nanoTime() aoqi@0: // these have the same type and signature aoqi@0: bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { aoqi@0: const TypeFunc* tf = OptoRuntime::void_long_Type(); aoqi@0: const TypePtr* no_memory_effects = NULL; aoqi@0: Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); aoqi@0: Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0)); aoqi@0: #ifdef ASSERT aoqi@0: Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1)); aoqi@0: assert(value_top == top(), "second value must be top"); aoqi@0: #endif aoqi@0: set_result(value); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------inline_native_currentThread------------------ aoqi@0: bool LibraryCallKit::inline_native_currentThread() { aoqi@0: Node* junk = NULL; aoqi@0: set_result(generate_current_thread(junk)); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------inline_native_isInterrupted------------------ aoqi@0: // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); aoqi@0: bool LibraryCallKit::inline_native_isInterrupted() { aoqi@0: // Add a fast path to t.isInterrupted(clear_int): aoqi@0: // (t == Thread.current() && aoqi@0: // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) aoqi@0: // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) aoqi@0: // So, in the common case that the interrupt bit is false, aoqi@0: // we avoid making a call into the VM. Even if the interrupt bit aoqi@0: // is true, if the clear_int argument is false, we avoid the VM call. aoqi@0: // However, if the receiver is not currentThread, we must call the VM, aoqi@0: // because there must be some locking done around the operation. aoqi@0: aoqi@0: // We only go to the fast case code if we pass two guards. aoqi@0: // Paths which do not pass are accumulated in the slow_region. aoqi@0: aoqi@0: enum { aoqi@0: no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted aoqi@0: no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int aoqi@0: slow_result_path = 3, // slow path: t.isInterrupted(clear_int) aoqi@0: PATH_LIMIT aoqi@0: }; aoqi@0: aoqi@0: // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag aoqi@0: // out of the function. aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: aoqi@0: RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT); aoqi@0: PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL); aoqi@0: aoqi@0: RegionNode* slow_region = new (C) RegionNode(1); aoqi@0: record_for_igvn(slow_region); aoqi@0: aoqi@0: // (a) Receiving thread must be the current thread. aoqi@0: Node* rec_thr = argument(0); aoqi@0: Node* tls_ptr = NULL; aoqi@0: Node* cur_thr = generate_current_thread(tls_ptr); aoqi@0: Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr)); aoqi@0: Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne)); aoqi@0: aoqi@0: generate_slow_guard(bol_thr, slow_region); aoqi@0: aoqi@0: // (b) Interrupt bit on TLS must be false. aoqi@0: Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); aoqi@0: Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); aoqi@0: p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); aoqi@0: aoqi@0: // Set the control input on the field _interrupted read to prevent it floating up. aoqi@0: Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); aoqi@0: Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0))); aoqi@0: Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne)); aoqi@0: aoqi@0: IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); aoqi@0: aoqi@0: // First fast path: if (!TLS._interrupted) return false; aoqi@0: Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit)); aoqi@0: result_rgn->init_req(no_int_result_path, false_bit); aoqi@0: result_val->init_req(no_int_result_path, intcon(0)); aoqi@0: aoqi@0: // drop through to next case aoqi@0: set_control( _gvn.transform(new (C) IfTrueNode(iff_bit))); aoqi@0: aoqi@0: #ifndef TARGET_OS_FAMILY_windows aoqi@0: // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. aoqi@0: Node* clr_arg = argument(1); aoqi@0: Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0))); aoqi@0: Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne)); aoqi@0: IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); aoqi@0: aoqi@0: // Second fast path: ... else if (!clear_int) return true; aoqi@0: Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg)); aoqi@0: result_rgn->init_req(no_clear_result_path, false_arg); aoqi@0: result_val->init_req(no_clear_result_path, intcon(1)); aoqi@0: aoqi@0: // drop through to next case aoqi@0: set_control( _gvn.transform(new (C) IfTrueNode(iff_arg))); aoqi@0: #else aoqi@0: // To return true on Windows you must read the _interrupted field aoqi@0: // and check the the event state i.e. take the slow path. aoqi@0: #endif // TARGET_OS_FAMILY_windows aoqi@0: aoqi@0: // (d) Otherwise, go to the slow path. aoqi@0: slow_region->add_req(control()); aoqi@0: set_control( _gvn.transform(slow_region)); aoqi@0: aoqi@0: if (stopped()) { aoqi@0: // There is no slow path. aoqi@0: result_rgn->init_req(slow_result_path, top()); aoqi@0: result_val->init_req(slow_result_path, top()); aoqi@0: } else { aoqi@0: // non-virtual because it is a private non-static aoqi@0: CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); aoqi@0: aoqi@0: Node* slow_val = set_results_for_java_call(slow_call); aoqi@0: // this->control() comes from set_results_for_java_call aoqi@0: aoqi@0: Node* fast_io = slow_call->in(TypeFunc::I_O); aoqi@0: Node* fast_mem = slow_call->in(TypeFunc::Memory); aoqi@0: aoqi@0: // These two phis are pre-filled with copies of of the fast IO and Memory aoqi@0: PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); aoqi@0: PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); aoqi@0: aoqi@0: result_rgn->init_req(slow_result_path, control()); aoqi@0: result_io ->init_req(slow_result_path, i_o()); aoqi@0: result_mem->init_req(slow_result_path, reset_memory()); aoqi@0: result_val->init_req(slow_result_path, slow_val); aoqi@0: aoqi@0: set_all_memory(_gvn.transform(result_mem)); aoqi@0: set_i_o( _gvn.transform(result_io)); aoqi@0: } aoqi@0: aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: set_result(result_rgn, result_val); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //---------------------------load_mirror_from_klass---------------------------- aoqi@0: // Given a klass oop, load its java mirror (a java.lang.Class oop). aoqi@0: Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { aoqi@0: Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); aoqi@0: return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); aoqi@0: } aoqi@0: aoqi@0: //-----------------------load_klass_from_mirror_common------------------------- aoqi@0: // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. aoqi@0: // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), aoqi@0: // and branch to the given path on the region. aoqi@0: // If never_see_null, take an uncommon trap on null, so we can optimistically aoqi@0: // compile for the non-null case. aoqi@0: // If the region is NULL, force never_see_null = true. aoqi@0: Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, aoqi@0: bool never_see_null, aoqi@0: RegionNode* region, aoqi@0: int null_path, aoqi@0: int offset) { aoqi@0: if (region == NULL) never_see_null = true; aoqi@0: Node* p = basic_plus_adr(mirror, offset); aoqi@0: const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; aoqi@0: Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); aoqi@0: Node* null_ctl = top(); aoqi@0: kls = null_check_oop(kls, &null_ctl, never_see_null); aoqi@0: if (region != NULL) { aoqi@0: // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). aoqi@0: region->init_req(null_path, null_ctl); aoqi@0: } else { aoqi@0: assert(null_ctl == top(), "no loose ends"); aoqi@0: } aoqi@0: return kls; aoqi@0: } aoqi@0: aoqi@0: //--------------------(inline_native_Class_query helpers)--------------------- aoqi@0: // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. aoqi@0: // Fall through if (mods & mask) == bits, take the guard otherwise. aoqi@0: Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { aoqi@0: // Branch around if the given klass has the given modifier bit set. aoqi@0: // Like generate_guard, adds a new path onto the region. aoqi@0: Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); aoqi@0: Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); aoqi@0: Node* mask = intcon(modifier_mask); aoqi@0: Node* bits = intcon(modifier_bits); aoqi@0: Node* mbit = _gvn.transform(new (C) AndINode(mods, mask)); aoqi@0: Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits)); aoqi@0: Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); aoqi@0: return generate_fair_guard(bol, region); aoqi@0: } aoqi@0: Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { aoqi@0: return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); aoqi@0: } aoqi@0: aoqi@0: //-------------------------inline_native_Class_query------------------- aoqi@0: bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { aoqi@0: const Type* return_type = TypeInt::BOOL; aoqi@0: Node* prim_return_value = top(); // what happens if it's a primitive class? aoqi@0: bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); aoqi@0: bool expect_prim = false; // most of these guys expect to work on refs aoqi@0: aoqi@0: enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; aoqi@0: aoqi@0: Node* mirror = argument(0); aoqi@0: Node* obj = top(); aoqi@0: aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_isInstance: aoqi@0: // nothing is an instance of a primitive type aoqi@0: prim_return_value = intcon(0); aoqi@0: obj = argument(1); aoqi@0: break; aoqi@0: case vmIntrinsics::_getModifiers: aoqi@0: prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); aoqi@0: assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); aoqi@0: return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); aoqi@0: break; aoqi@0: case vmIntrinsics::_isInterface: aoqi@0: prim_return_value = intcon(0); aoqi@0: break; aoqi@0: case vmIntrinsics::_isArray: aoqi@0: prim_return_value = intcon(0); aoqi@0: expect_prim = true; // cf. ObjectStreamClass.getClassSignature aoqi@0: break; aoqi@0: case vmIntrinsics::_isPrimitive: aoqi@0: prim_return_value = intcon(1); aoqi@0: expect_prim = true; // obviously aoqi@0: break; aoqi@0: case vmIntrinsics::_getSuperclass: aoqi@0: prim_return_value = null(); aoqi@0: return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); aoqi@0: break; aoqi@0: case vmIntrinsics::_getComponentType: aoqi@0: prim_return_value = null(); aoqi@0: return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); aoqi@0: break; aoqi@0: case vmIntrinsics::_getClassAccessFlags: aoqi@0: prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); aoqi@0: return_type = TypeInt::INT; // not bool! 6297094 aoqi@0: break; aoqi@0: default: aoqi@0: fatal_unexpected_iid(id); aoqi@0: break; aoqi@0: } aoqi@0: aoqi@0: const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); aoqi@0: if (mirror_con == NULL) return false; // cannot happen? aoqi@0: aoqi@0: #ifndef PRODUCT aoqi@0: if (C->print_intrinsics() || C->print_inlining()) { aoqi@0: ciType* k = mirror_con->java_mirror_type(); aoqi@0: if (k) { aoqi@0: tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); aoqi@0: k->print_name(); aoqi@0: tty->cr(); aoqi@0: } aoqi@0: } aoqi@0: #endif aoqi@0: aoqi@0: // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). aoqi@0: RegionNode* region = new (C) RegionNode(PATH_LIMIT); aoqi@0: record_for_igvn(region); aoqi@0: PhiNode* phi = new (C) PhiNode(region, return_type); aoqi@0: aoqi@0: // The mirror will never be null of Reflection.getClassAccessFlags, however aoqi@0: // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE aoqi@0: // if it is. See bug 4774291. aoqi@0: aoqi@0: // For Reflection.getClassAccessFlags(), the null check occurs in aoqi@0: // the wrong place; see inline_unsafe_access(), above, for a similar aoqi@0: // situation. aoqi@0: mirror = null_check(mirror); aoqi@0: // If mirror or obj is dead, only null-path is taken. aoqi@0: if (stopped()) return true; aoqi@0: aoqi@0: if (expect_prim) never_see_null = false; // expect nulls (meaning prims) aoqi@0: aoqi@0: // Now load the mirror's klass metaobject, and null-check it. aoqi@0: // Side-effects region with the control path if the klass is null. aoqi@0: Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); aoqi@0: // If kls is null, we have a primitive mirror. aoqi@0: phi->init_req(_prim_path, prim_return_value); aoqi@0: if (stopped()) { set_result(region, phi); return true; } aoqi@0: bool safe_for_replace = (region->in(_prim_path) == top()); aoqi@0: aoqi@0: Node* p; // handy temp aoqi@0: Node* null_ctl; aoqi@0: aoqi@0: // Now that we have the non-null klass, we can perform the real query. aoqi@0: // For constant classes, the query will constant-fold in LoadNode::Value. aoqi@0: Node* query_value = top(); aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_isInstance: aoqi@0: // nothing is an instance of a primitive type aoqi@0: query_value = gen_instanceof(obj, kls, safe_for_replace); aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_getModifiers: aoqi@0: p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); aoqi@0: query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_isInterface: aoqi@0: // (To verify this code sequence, check the asserts in JVM_IsInterface.) aoqi@0: if (generate_interface_guard(kls, region) != NULL) aoqi@0: // A guard was added. If the guard is taken, it was an interface. aoqi@0: phi->add_req(intcon(1)); aoqi@0: // If we fall through, it's a plain class. aoqi@0: query_value = intcon(0); aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_isArray: aoqi@0: // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) aoqi@0: if (generate_array_guard(kls, region) != NULL) aoqi@0: // A guard was added. If the guard is taken, it was an array. aoqi@0: phi->add_req(intcon(1)); aoqi@0: // If we fall through, it's a plain class. aoqi@0: query_value = intcon(0); aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_isPrimitive: aoqi@0: query_value = intcon(0); // "normal" path produces false aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_getSuperclass: aoqi@0: // The rules here are somewhat unfortunate, but we can still do better aoqi@0: // with random logic than with a JNI call. aoqi@0: // Interfaces store null or Object as _super, but must report null. aoqi@0: // Arrays store an intermediate super as _super, but must report Object. aoqi@0: // Other types can report the actual _super. aoqi@0: // (To verify this code sequence, check the asserts in JVM_IsInterface.) aoqi@0: if (generate_interface_guard(kls, region) != NULL) aoqi@0: // A guard was added. If the guard is taken, it was an interface. aoqi@0: phi->add_req(null()); aoqi@0: if (generate_array_guard(kls, region) != NULL) aoqi@0: // A guard was added. If the guard is taken, it was an array. aoqi@0: phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); aoqi@0: // If we fall through, it's a plain class. Get its _super. aoqi@0: p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); aoqi@0: kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); aoqi@0: null_ctl = top(); aoqi@0: kls = null_check_oop(kls, &null_ctl); aoqi@0: if (null_ctl != top()) { aoqi@0: // If the guard is taken, Object.superClass is null (both klass and mirror). aoqi@0: region->add_req(null_ctl); aoqi@0: phi ->add_req(null()); aoqi@0: } aoqi@0: if (!stopped()) { aoqi@0: query_value = load_mirror_from_klass(kls); aoqi@0: } aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_getComponentType: aoqi@0: if (generate_array_guard(kls, region) != NULL) { aoqi@0: // Be sure to pin the oop load to the guard edge just created: aoqi@0: Node* is_array_ctrl = region->in(region->req()-1); aoqi@0: Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset())); aoqi@0: Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); aoqi@0: phi->add_req(cmo); aoqi@0: } aoqi@0: query_value = null(); // non-array case is null aoqi@0: break; aoqi@0: aoqi@0: case vmIntrinsics::_getClassAccessFlags: aoqi@0: p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); aoqi@0: query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); aoqi@0: break; aoqi@0: aoqi@0: default: aoqi@0: fatal_unexpected_iid(id); aoqi@0: break; aoqi@0: } aoqi@0: aoqi@0: // Fall-through is the normal case of a query to a real class. aoqi@0: phi->init_req(1, query_value); aoqi@0: region->init_req(1, control()); aoqi@0: aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: set_result(region, phi); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //--------------------------inline_native_subtype_check------------------------ aoqi@0: // This intrinsic takes the JNI calls out of the heart of aoqi@0: // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. aoqi@0: bool LibraryCallKit::inline_native_subtype_check() { aoqi@0: // Pull both arguments off the stack. aoqi@0: Node* args[2]; // two java.lang.Class mirrors: superc, subc aoqi@0: args[0] = argument(0); aoqi@0: args[1] = argument(1); aoqi@0: Node* klasses[2]; // corresponding Klasses: superk, subk aoqi@0: klasses[0] = klasses[1] = top(); aoqi@0: aoqi@0: enum { aoqi@0: // A full decision tree on {superc is prim, subc is prim}: aoqi@0: _prim_0_path = 1, // {P,N} => false aoqi@0: // {P,P} & superc!=subc => false aoqi@0: _prim_same_path, // {P,P} & superc==subc => true aoqi@0: _prim_1_path, // {N,P} => false aoqi@0: _ref_subtype_path, // {N,N} & subtype check wins => true aoqi@0: _both_ref_path, // {N,N} & subtype check loses => false aoqi@0: PATH_LIMIT aoqi@0: }; aoqi@0: aoqi@0: RegionNode* region = new (C) RegionNode(PATH_LIMIT); aoqi@0: Node* phi = new (C) PhiNode(region, TypeInt::BOOL); aoqi@0: record_for_igvn(region); aoqi@0: aoqi@0: const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads aoqi@0: const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; aoqi@0: int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); aoqi@0: aoqi@0: // First null-check both mirrors and load each mirror's klass metaobject. aoqi@0: int which_arg; aoqi@0: for (which_arg = 0; which_arg <= 1; which_arg++) { aoqi@0: Node* arg = args[which_arg]; aoqi@0: arg = null_check(arg); aoqi@0: if (stopped()) break; aoqi@0: args[which_arg] = arg; aoqi@0: aoqi@0: Node* p = basic_plus_adr(arg, class_klass_offset); aoqi@0: Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); aoqi@0: klasses[which_arg] = _gvn.transform(kls); aoqi@0: } aoqi@0: aoqi@0: // Having loaded both klasses, test each for null. aoqi@0: bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); aoqi@0: for (which_arg = 0; which_arg <= 1; which_arg++) { aoqi@0: Node* kls = klasses[which_arg]; aoqi@0: Node* null_ctl = top(); aoqi@0: kls = null_check_oop(kls, &null_ctl, never_see_null); aoqi@0: int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); aoqi@0: region->init_req(prim_path, null_ctl); aoqi@0: if (stopped()) break; aoqi@0: klasses[which_arg] = kls; aoqi@0: } aoqi@0: aoqi@0: if (!stopped()) { aoqi@0: // now we have two reference types, in klasses[0..1] aoqi@0: Node* subk = klasses[1]; // the argument to isAssignableFrom aoqi@0: Node* superk = klasses[0]; // the receiver aoqi@0: region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); aoqi@0: // now we have a successful reference subtype check aoqi@0: region->set_req(_ref_subtype_path, control()); aoqi@0: } aoqi@0: aoqi@0: // If both operands are primitive (both klasses null), then aoqi@0: // we must return true when they are identical primitives. aoqi@0: // It is convenient to test this after the first null klass check. aoqi@0: set_control(region->in(_prim_0_path)); // go back to first null check aoqi@0: if (!stopped()) { aoqi@0: // Since superc is primitive, make a guard for the superc==subc case. aoqi@0: Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1])); aoqi@0: Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq)); aoqi@0: generate_guard(bol_eq, region, PROB_FAIR); aoqi@0: if (region->req() == PATH_LIMIT+1) { aoqi@0: // A guard was added. If the added guard is taken, superc==subc. aoqi@0: region->swap_edges(PATH_LIMIT, _prim_same_path); aoqi@0: region->del_req(PATH_LIMIT); aoqi@0: } aoqi@0: region->set_req(_prim_0_path, control()); // Not equal after all. aoqi@0: } aoqi@0: aoqi@0: // these are the only paths that produce 'true': aoqi@0: phi->set_req(_prim_same_path, intcon(1)); aoqi@0: phi->set_req(_ref_subtype_path, intcon(1)); aoqi@0: aoqi@0: // pull together the cases: aoqi@0: assert(region->req() == PATH_LIMIT, "sane region"); aoqi@0: for (uint i = 1; i < region->req(); i++) { aoqi@0: Node* ctl = region->in(i); aoqi@0: if (ctl == NULL || ctl == top()) { aoqi@0: region->set_req(i, top()); aoqi@0: phi ->set_req(i, top()); aoqi@0: } else if (phi->in(i) == NULL) { aoqi@0: phi->set_req(i, intcon(0)); // all other paths produce 'false' aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: set_control(_gvn.transform(region)); aoqi@0: set_result(_gvn.transform(phi)); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //---------------------generate_array_guard_common------------------------ aoqi@0: Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, aoqi@0: bool obj_array, bool not_array) { aoqi@0: // If obj_array/non_array==false/false: aoqi@0: // Branch around if the given klass is in fact an array (either obj or prim). aoqi@0: // If obj_array/non_array==false/true: aoqi@0: // Branch around if the given klass is not an array klass of any kind. aoqi@0: // If obj_array/non_array==true/true: aoqi@0: // Branch around if the kls is not an oop array (kls is int[], String, etc.) aoqi@0: // If obj_array/non_array==true/false: aoqi@0: // Branch around if the kls is an oop array (Object[] or subtype) aoqi@0: // aoqi@0: // Like generate_guard, adds a new path onto the region. aoqi@0: jint layout_con = 0; aoqi@0: Node* layout_val = get_layout_helper(kls, layout_con); aoqi@0: if (layout_val == NULL) { aoqi@0: bool query = (obj_array aoqi@0: ? Klass::layout_helper_is_objArray(layout_con) aoqi@0: : Klass::layout_helper_is_array(layout_con)); aoqi@0: if (query == not_array) { aoqi@0: return NULL; // never a branch aoqi@0: } else { // always a branch aoqi@0: Node* always_branch = control(); aoqi@0: if (region != NULL) aoqi@0: region->add_req(always_branch); aoqi@0: set_control(top()); aoqi@0: return always_branch; aoqi@0: } aoqi@0: } aoqi@0: // Now test the correct condition. aoqi@0: jint nval = (obj_array aoqi@0: ? ((jint)Klass::_lh_array_tag_type_value aoqi@0: << Klass::_lh_array_tag_shift) aoqi@0: : Klass::_lh_neutral_value); aoqi@0: Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval))); aoqi@0: BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array aoqi@0: // invert the test if we are looking for a non-array aoqi@0: if (not_array) btest = BoolTest(btest).negate(); aoqi@0: Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest)); aoqi@0: return generate_fair_guard(bol, region); aoqi@0: } aoqi@0: aoqi@0: aoqi@0: //-----------------------inline_native_newArray-------------------------- aoqi@0: // private static native Object java.lang.reflect.newArray(Class componentType, int length); aoqi@0: bool LibraryCallKit::inline_native_newArray() { aoqi@0: Node* mirror = argument(0); aoqi@0: Node* count_val = argument(1); aoqi@0: aoqi@0: mirror = null_check(mirror); aoqi@0: // If mirror or obj is dead, only null-path is taken. aoqi@0: if (stopped()) return true; aoqi@0: aoqi@0: enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; aoqi@0: RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); aoqi@0: PhiNode* result_val = new(C) PhiNode(result_reg, aoqi@0: TypeInstPtr::NOTNULL); aoqi@0: PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); aoqi@0: PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, aoqi@0: TypePtr::BOTTOM); aoqi@0: aoqi@0: bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); aoqi@0: Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, aoqi@0: result_reg, _slow_path); aoqi@0: Node* normal_ctl = control(); aoqi@0: Node* no_array_ctl = result_reg->in(_slow_path); aoqi@0: aoqi@0: // Generate code for the slow case. We make a call to newArray(). aoqi@0: set_control(no_array_ctl); aoqi@0: if (!stopped()) { aoqi@0: // Either the input type is void.class, or else the aoqi@0: // array klass has not yet been cached. Either the aoqi@0: // ensuing call will throw an exception, or else it aoqi@0: // will cache the array klass for next time. aoqi@0: PreserveJVMState pjvms(this); aoqi@0: CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); aoqi@0: Node* slow_result = set_results_for_java_call(slow_call); aoqi@0: // this->control() comes from set_results_for_java_call aoqi@0: result_reg->set_req(_slow_path, control()); aoqi@0: result_val->set_req(_slow_path, slow_result); aoqi@0: result_io ->set_req(_slow_path, i_o()); aoqi@0: result_mem->set_req(_slow_path, reset_memory()); aoqi@0: } aoqi@0: aoqi@0: set_control(normal_ctl); aoqi@0: if (!stopped()) { aoqi@0: // Normal case: The array type has been cached in the java.lang.Class. aoqi@0: // The following call works fine even if the array type is polymorphic. aoqi@0: // It could be a dynamic mix of int[], boolean[], Object[], etc. aoqi@0: Node* obj = new_array(klass_node, count_val, 0); // no arguments to push aoqi@0: result_reg->init_req(_normal_path, control()); aoqi@0: result_val->init_req(_normal_path, obj); aoqi@0: result_io ->init_req(_normal_path, i_o()); aoqi@0: result_mem->init_req(_normal_path, reset_memory()); aoqi@0: } aoqi@0: aoqi@0: // Return the combined state. aoqi@0: set_i_o( _gvn.transform(result_io) ); aoqi@0: set_all_memory( _gvn.transform(result_mem)); aoqi@0: aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: set_result(result_reg, result_val); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //----------------------inline_native_getLength-------------------------- aoqi@0: // public static native int java.lang.reflect.Array.getLength(Object array); aoqi@0: bool LibraryCallKit::inline_native_getLength() { aoqi@0: if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; aoqi@0: aoqi@0: Node* array = null_check(argument(0)); aoqi@0: // If array is dead, only null-path is taken. aoqi@0: if (stopped()) return true; aoqi@0: aoqi@0: // Deoptimize if it is a non-array. aoqi@0: Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); aoqi@0: aoqi@0: if (non_array != NULL) { aoqi@0: PreserveJVMState pjvms(this); aoqi@0: set_control(non_array); aoqi@0: uncommon_trap(Deoptimization::Reason_intrinsic, aoqi@0: Deoptimization::Action_maybe_recompile); aoqi@0: } aoqi@0: aoqi@0: // If control is dead, only non-array-path is taken. aoqi@0: if (stopped()) return true; aoqi@0: aoqi@0: // The works fine even if the array type is polymorphic. aoqi@0: // It could be a dynamic mix of int[], boolean[], Object[], etc. aoqi@0: Node* result = load_array_length(array); aoqi@0: aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: set_result(result); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------inline_array_copyOf---------------------------- aoqi@0: // public static T[] java.util.Arrays.copyOf( U[] original, int newLength, Class newType); aoqi@0: // public static T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class newType); aoqi@0: bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { aoqi@0: if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; aoqi@0: aoqi@0: // Get the arguments. aoqi@0: Node* original = argument(0); aoqi@0: Node* start = is_copyOfRange? argument(1): intcon(0); aoqi@0: Node* end = is_copyOfRange? argument(2): argument(1); aoqi@0: Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); aoqi@0: aoqi@0: Node* newcopy; aoqi@0: aoqi@0: // Set the original stack and the reexecute bit for the interpreter to reexecute aoqi@0: // the bytecode that invokes Arrays.copyOf if deoptimization happens. aoqi@0: { PreserveReexecuteState preexecs(this); aoqi@0: jvms()->set_should_reexecute(true); aoqi@0: aoqi@0: array_type_mirror = null_check(array_type_mirror); aoqi@0: original = null_check(original); aoqi@0: aoqi@0: // Check if a null path was taken unconditionally. aoqi@0: if (stopped()) return true; aoqi@0: aoqi@0: Node* orig_length = load_array_length(original); aoqi@0: aoqi@0: Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); aoqi@0: klass_node = null_check(klass_node); aoqi@0: aoqi@0: RegionNode* bailout = new (C) RegionNode(1); aoqi@0: record_for_igvn(bailout); aoqi@0: aoqi@0: // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. aoqi@0: // Bail out if that is so. aoqi@0: Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); aoqi@0: if (not_objArray != NULL) { aoqi@0: // Improve the klass node's type from the new optimistic assumption: aoqi@0: ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); aoqi@0: const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); aoqi@0: Node* cast = new (C) CastPPNode(klass_node, akls); aoqi@0: cast->init_req(0, control()); aoqi@0: klass_node = _gvn.transform(cast); aoqi@0: } aoqi@0: aoqi@0: // Bail out if either start or end is negative. aoqi@0: generate_negative_guard(start, bailout, &start); aoqi@0: generate_negative_guard(end, bailout, &end); aoqi@0: aoqi@0: Node* length = end; aoqi@0: if (_gvn.type(start) != TypeInt::ZERO) { aoqi@0: length = _gvn.transform(new (C) SubINode(end, start)); aoqi@0: } aoqi@0: aoqi@0: // Bail out if length is negative. aoqi@0: // Without this the new_array would throw aoqi@0: // NegativeArraySizeException but IllegalArgumentException is what aoqi@0: // should be thrown aoqi@0: generate_negative_guard(length, bailout, &length); aoqi@0: aoqi@0: if (bailout->req() > 1) { aoqi@0: PreserveJVMState pjvms(this); aoqi@0: set_control(_gvn.transform(bailout)); aoqi@0: uncommon_trap(Deoptimization::Reason_intrinsic, aoqi@0: Deoptimization::Action_maybe_recompile); aoqi@0: } aoqi@0: aoqi@0: if (!stopped()) { aoqi@0: // How many elements will we copy from the original? aoqi@0: // The answer is MinI(orig_length - start, length). aoqi@0: Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start)); aoqi@0: Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); aoqi@0: aoqi@0: newcopy = new_array(klass_node, length, 0); // no argments to push aoqi@0: aoqi@0: // Generate a direct call to the right arraycopy function(s). aoqi@0: // We know the copy is disjoint but we might not know if the aoqi@0: // oop stores need checking. aoqi@0: // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). aoqi@0: // This will fail a store-check if x contains any non-nulls. aoqi@0: bool disjoint_bases = true; aoqi@0: // if start > orig_length then the length of the copy may be aoqi@0: // negative. aoqi@0: bool length_never_negative = !is_copyOfRange; aoqi@0: generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, aoqi@0: original, start, newcopy, intcon(0), moved, aoqi@0: disjoint_bases, length_never_negative); aoqi@0: } aoqi@0: } // original reexecute is set back here aoqi@0: aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: if (!stopped()) { aoqi@0: set_result(newcopy); aoqi@0: } aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: aoqi@0: //----------------------generate_virtual_guard--------------------------- aoqi@0: // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. aoqi@0: Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, aoqi@0: RegionNode* slow_region) { aoqi@0: ciMethod* method = callee(); aoqi@0: int vtable_index = method->vtable_index(); aoqi@0: assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, aoqi@0: err_msg_res("bad index %d", vtable_index)); aoqi@0: // Get the Method* out of the appropriate vtable entry. aoqi@0: int entry_offset = (InstanceKlass::vtable_start_offset() + aoqi@0: vtable_index*vtableEntry::size()) * wordSize + aoqi@0: vtableEntry::method_offset_in_bytes(); aoqi@0: Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); aoqi@0: Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); aoqi@0: aoqi@0: // Compare the target method with the expected method (e.g., Object.hashCode). aoqi@0: const TypePtr* native_call_addr = TypeMetadataPtr::make(method); aoqi@0: aoqi@0: Node* native_call = makecon(native_call_addr); aoqi@0: Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call)); aoqi@0: Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne)); aoqi@0: aoqi@0: return generate_slow_guard(test_native, slow_region); aoqi@0: } aoqi@0: aoqi@0: //-----------------------generate_method_call---------------------------- aoqi@0: // Use generate_method_call to make a slow-call to the real aoqi@0: // method if the fast path fails. An alternative would be to aoqi@0: // use a stub like OptoRuntime::slow_arraycopy_Java. aoqi@0: // This only works for expanding the current library call, aoqi@0: // not another intrinsic. (E.g., don't use this for making an aoqi@0: // arraycopy call inside of the copyOf intrinsic.) aoqi@0: CallJavaNode* aoqi@0: LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { aoqi@0: // When compiling the intrinsic method itself, do not use this technique. aoqi@0: guarantee(callee() != C->method(), "cannot make slow-call to self"); aoqi@0: aoqi@0: ciMethod* method = callee(); aoqi@0: // ensure the JVMS we have will be correct for this call aoqi@0: guarantee(method_id == method->intrinsic_id(), "must match"); aoqi@0: aoqi@0: const TypeFunc* tf = TypeFunc::make(method); aoqi@0: CallJavaNode* slow_call; aoqi@0: if (is_static) { aoqi@0: assert(!is_virtual, ""); aoqi@0: slow_call = new(C) CallStaticJavaNode(C, tf, aoqi@0: SharedRuntime::get_resolve_static_call_stub(), aoqi@0: method, bci()); aoqi@0: } else if (is_virtual) { aoqi@0: null_check_receiver(); aoqi@0: int vtable_index = Method::invalid_vtable_index; aoqi@0: if (UseInlineCaches) { aoqi@0: // Suppress the vtable call aoqi@0: } else { aoqi@0: // hashCode and clone are not a miranda methods, aoqi@0: // so the vtable index is fixed. aoqi@0: // No need to use the linkResolver to get it. aoqi@0: vtable_index = method->vtable_index(); aoqi@0: assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, aoqi@0: err_msg_res("bad index %d", vtable_index)); aoqi@0: } aoqi@0: slow_call = new(C) CallDynamicJavaNode(tf, aoqi@0: SharedRuntime::get_resolve_virtual_call_stub(), aoqi@0: method, vtable_index, bci()); aoqi@0: } else { // neither virtual nor static: opt_virtual aoqi@0: null_check_receiver(); aoqi@0: slow_call = new(C) CallStaticJavaNode(C, tf, aoqi@0: SharedRuntime::get_resolve_opt_virtual_call_stub(), aoqi@0: method, bci()); aoqi@0: slow_call->set_optimized_virtual(true); aoqi@0: } aoqi@0: set_arguments_for_java_call(slow_call); aoqi@0: set_edges_for_java_call(slow_call); aoqi@0: return slow_call; aoqi@0: } aoqi@0: aoqi@0: aoqi@0: /** aoqi@0: * Build special case code for calls to hashCode on an object. This call may aoqi@0: * be virtual (invokevirtual) or bound (invokespecial). For each case we generate aoqi@0: * slightly different code. aoqi@0: */ aoqi@0: bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { aoqi@0: assert(is_static == callee()->is_static(), "correct intrinsic selection"); aoqi@0: assert(!(is_virtual && is_static), "either virtual, special, or static"); aoqi@0: aoqi@0: enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; aoqi@0: aoqi@0: RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); aoqi@0: PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT); aoqi@0: PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); aoqi@0: PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); aoqi@0: Node* obj = NULL; aoqi@0: if (!is_static) { aoqi@0: // Check for hashing null object aoqi@0: obj = null_check_receiver(); aoqi@0: if (stopped()) return true; // unconditionally null aoqi@0: result_reg->init_req(_null_path, top()); aoqi@0: result_val->init_req(_null_path, top()); aoqi@0: } else { aoqi@0: // Do a null check, and return zero if null. aoqi@0: // System.identityHashCode(null) == 0 aoqi@0: obj = argument(0); aoqi@0: Node* null_ctl = top(); aoqi@0: obj = null_check_oop(obj, &null_ctl); aoqi@0: result_reg->init_req(_null_path, null_ctl); aoqi@0: result_val->init_req(_null_path, _gvn.intcon(0)); aoqi@0: } aoqi@0: aoqi@0: // Unconditionally null? Then return right away. aoqi@0: if (stopped()) { aoqi@0: set_control( result_reg->in(_null_path)); aoqi@0: if (!stopped()) aoqi@0: set_result(result_val->in(_null_path)); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: // We only go to the fast case code if we pass a number of guards. The aoqi@0: // paths which do not pass are accumulated in the slow_region. aoqi@0: RegionNode* slow_region = new (C) RegionNode(1); aoqi@0: record_for_igvn(slow_region); aoqi@0: aoqi@0: // If this is a virtual call, we generate a funny guard. We pull out aoqi@0: // the vtable entry corresponding to hashCode() from the target object. aoqi@0: // If the target method which we are calling happens to be the native aoqi@0: // Object hashCode() method, we pass the guard. We do not need this aoqi@0: // guard for non-virtual calls -- the caller is known to be the native aoqi@0: // Object hashCode(). aoqi@0: if (is_virtual) { aoqi@0: // After null check, get the object's klass. aoqi@0: Node* obj_klass = load_object_klass(obj); aoqi@0: generate_virtual_guard(obj_klass, slow_region); aoqi@0: } aoqi@0: aoqi@0: // Get the header out of the object, use LoadMarkNode when available aoqi@0: Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); aoqi@0: // The control of the load must be NULL. Otherwise, the load can move before aoqi@0: // the null check after castPP removal. aoqi@0: Node* no_ctrl = NULL; aoqi@0: Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); aoqi@0: aoqi@0: // Test the header to see if it is unlocked. aoqi@0: Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); aoqi@0: Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask)); aoqi@0: Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); aoqi@0: Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val)); aoqi@0: Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne)); aoqi@0: aoqi@0: generate_slow_guard(test_unlocked, slow_region); aoqi@0: aoqi@0: // Get the hash value and check to see that it has been properly assigned. aoqi@0: // We depend on hash_mask being at most 32 bits and avoid the use of aoqi@0: // hash_mask_in_place because it could be larger than 32 bits in a 64-bit aoqi@0: // vm: see markOop.hpp. aoqi@0: Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask); aoqi@0: Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift); aoqi@0: Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift)); aoqi@0: // This hack lets the hash bits live anywhere in the mark object now, as long aoqi@0: // as the shift drops the relevant bits into the low 32 bits. Note that aoqi@0: // Java spec says that HashCode is an int so there's no point in capturing aoqi@0: // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). aoqi@0: hshifted_header = ConvX2I(hshifted_header); aoqi@0: Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask)); aoqi@0: aoqi@0: Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash); aoqi@0: Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val)); aoqi@0: Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq)); aoqi@0: aoqi@0: generate_slow_guard(test_assigned, slow_region); aoqi@0: aoqi@0: Node* init_mem = reset_memory(); aoqi@0: // fill in the rest of the null path: aoqi@0: result_io ->init_req(_null_path, i_o()); aoqi@0: result_mem->init_req(_null_path, init_mem); aoqi@0: aoqi@0: result_val->init_req(_fast_path, hash_val); aoqi@0: result_reg->init_req(_fast_path, control()); aoqi@0: result_io ->init_req(_fast_path, i_o()); aoqi@0: result_mem->init_req(_fast_path, init_mem); aoqi@0: aoqi@0: // Generate code for the slow case. We make a call to hashCode(). aoqi@0: set_control(_gvn.transform(slow_region)); aoqi@0: if (!stopped()) { aoqi@0: // No need for PreserveJVMState, because we're using up the present state. aoqi@0: set_all_memory(init_mem); aoqi@0: vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; aoqi@0: CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); aoqi@0: Node* slow_result = set_results_for_java_call(slow_call); aoqi@0: // this->control() comes from set_results_for_java_call aoqi@0: result_reg->init_req(_slow_path, control()); aoqi@0: result_val->init_req(_slow_path, slow_result); aoqi@0: result_io ->set_req(_slow_path, i_o()); aoqi@0: result_mem ->set_req(_slow_path, reset_memory()); aoqi@0: } aoqi@0: aoqi@0: // Return the combined state. aoqi@0: set_i_o( _gvn.transform(result_io) ); aoqi@0: set_all_memory( _gvn.transform(result_mem)); aoqi@0: aoqi@0: set_result(result_reg, result_val); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //---------------------------inline_native_getClass---------------------------- aoqi@0: // public final native Class java.lang.Object.getClass(); aoqi@0: // aoqi@0: // Build special case code for calls to getClass on an object. aoqi@0: bool LibraryCallKit::inline_native_getClass() { aoqi@0: Node* obj = null_check_receiver(); aoqi@0: if (stopped()) return true; aoqi@0: set_result(load_mirror_from_klass(load_object_klass(obj))); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //-----------------inline_native_Reflection_getCallerClass--------------------- aoqi@0: // public static native Class sun.reflect.Reflection.getCallerClass(); aoqi@0: // aoqi@0: // In the presence of deep enough inlining, getCallerClass() becomes a no-op. aoqi@0: // aoqi@0: // NOTE: This code must perform the same logic as JVM_GetCallerClass aoqi@0: // in that it must skip particular security frames and checks for aoqi@0: // caller sensitive methods. aoqi@0: bool LibraryCallKit::inline_native_Reflection_getCallerClass() { aoqi@0: #ifndef PRODUCT aoqi@0: if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { aoqi@0: tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); aoqi@0: } aoqi@0: #endif aoqi@0: aoqi@0: if (!jvms()->has_method()) { aoqi@0: #ifndef PRODUCT aoqi@0: if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { aoqi@0: tty->print_cr(" Bailing out because intrinsic was inlined at top level"); aoqi@0: } aoqi@0: #endif aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: // Walk back up the JVM state to find the caller at the required aoqi@0: // depth. aoqi@0: JVMState* caller_jvms = jvms(); aoqi@0: aoqi@0: // Cf. JVM_GetCallerClass aoqi@0: // NOTE: Start the loop at depth 1 because the current JVM state does aoqi@0: // not include the Reflection.getCallerClass() frame. aoqi@0: for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { aoqi@0: ciMethod* m = caller_jvms->method(); aoqi@0: switch (n) { aoqi@0: case 0: aoqi@0: fatal("current JVM state does not include the Reflection.getCallerClass frame"); aoqi@0: break; aoqi@0: case 1: aoqi@0: // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). aoqi@0: if (!m->caller_sensitive()) { aoqi@0: #ifndef PRODUCT aoqi@0: if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { aoqi@0: tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); aoqi@0: } aoqi@0: #endif aoqi@0: return false; // bail-out; let JVM_GetCallerClass do the work aoqi@0: } aoqi@0: break; aoqi@0: default: aoqi@0: if (!m->is_ignored_by_security_stack_walk()) { aoqi@0: // We have reached the desired frame; return the holder class. aoqi@0: // Acquire method holder as java.lang.Class and push as constant. aoqi@0: ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); aoqi@0: ciInstance* caller_mirror = caller_klass->java_mirror(); aoqi@0: set_result(makecon(TypeInstPtr::make(caller_mirror))); aoqi@0: aoqi@0: #ifndef PRODUCT aoqi@0: if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { aoqi@0: tty->print_cr(" Succeeded: caller = %d) %s.%s, JVMS depth = %d", n, caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), jvms()->depth()); aoqi@0: tty->print_cr(" JVM state at this point:"); aoqi@0: for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { aoqi@0: ciMethod* m = jvms()->of_depth(i)->method(); aoqi@0: tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); aoqi@0: } aoqi@0: } aoqi@0: #endif aoqi@0: return true; aoqi@0: } aoqi@0: break; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: #ifndef PRODUCT aoqi@0: if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { aoqi@0: tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); aoqi@0: tty->print_cr(" JVM state at this point:"); aoqi@0: for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { aoqi@0: ciMethod* m = jvms()->of_depth(i)->method(); aoqi@0: tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); aoqi@0: } aoqi@0: } aoqi@0: #endif aoqi@0: aoqi@0: return false; // bail-out; let JVM_GetCallerClass do the work aoqi@0: } aoqi@0: aoqi@0: bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { aoqi@0: Node* arg = argument(0); aoqi@0: Node* result; aoqi@0: aoqi@0: switch (id) { aoqi@0: case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break; aoqi@0: case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break; aoqi@0: case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break; aoqi@0: case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break; aoqi@0: aoqi@0: case vmIntrinsics::_doubleToLongBits: { aoqi@0: // two paths (plus control) merge in a wood aoqi@0: RegionNode *r = new (C) RegionNode(3); aoqi@0: Node *phi = new (C) PhiNode(r, TypeLong::LONG); aoqi@0: aoqi@0: Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg)); aoqi@0: // Build the boolean node aoqi@0: Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); aoqi@0: aoqi@0: // Branch either way. aoqi@0: // NaN case is less traveled, which makes all the difference. aoqi@0: IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); aoqi@0: Node *opt_isnan = _gvn.transform(ifisnan); aoqi@0: assert( opt_isnan->is_If(), "Expect an IfNode"); aoqi@0: IfNode *opt_ifisnan = (IfNode*)opt_isnan; aoqi@0: Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); aoqi@0: aoqi@0: set_control(iftrue); aoqi@0: aoqi@0: static const jlong nan_bits = CONST64(0x7ff8000000000000); aoqi@0: Node *slow_result = longcon(nan_bits); // return NaN aoqi@0: phi->init_req(1, _gvn.transform( slow_result )); aoqi@0: r->init_req(1, iftrue); aoqi@0: aoqi@0: // Else fall through aoqi@0: Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); aoqi@0: set_control(iffalse); aoqi@0: aoqi@0: phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg))); aoqi@0: r->init_req(2, iffalse); aoqi@0: aoqi@0: // Post merge aoqi@0: set_control(_gvn.transform(r)); aoqi@0: record_for_igvn(r); aoqi@0: aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: result = phi; aoqi@0: assert(result->bottom_type()->isa_long(), "must be"); aoqi@0: break; aoqi@0: } aoqi@0: aoqi@0: case vmIntrinsics::_floatToIntBits: { aoqi@0: // two paths (plus control) merge in a wood aoqi@0: RegionNode *r = new (C) RegionNode(3); aoqi@0: Node *phi = new (C) PhiNode(r, TypeInt::INT); aoqi@0: aoqi@0: Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg)); aoqi@0: // Build the boolean node aoqi@0: Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); aoqi@0: aoqi@0: // Branch either way. aoqi@0: // NaN case is less traveled, which makes all the difference. aoqi@0: IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); aoqi@0: Node *opt_isnan = _gvn.transform(ifisnan); aoqi@0: assert( opt_isnan->is_If(), "Expect an IfNode"); aoqi@0: IfNode *opt_ifisnan = (IfNode*)opt_isnan; aoqi@0: Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); aoqi@0: aoqi@0: set_control(iftrue); aoqi@0: aoqi@0: static const jint nan_bits = 0x7fc00000; aoqi@0: Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN aoqi@0: phi->init_req(1, _gvn.transform( slow_result )); aoqi@0: r->init_req(1, iftrue); aoqi@0: aoqi@0: // Else fall through aoqi@0: Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); aoqi@0: set_control(iffalse); aoqi@0: aoqi@0: phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg))); aoqi@0: r->init_req(2, iffalse); aoqi@0: aoqi@0: // Post merge aoqi@0: set_control(_gvn.transform(r)); aoqi@0: record_for_igvn(r); aoqi@0: aoqi@0: C->set_has_split_ifs(true); // Has chance for split-if optimization aoqi@0: result = phi; aoqi@0: assert(result->bottom_type()->isa_int(), "must be"); aoqi@0: break; aoqi@0: } aoqi@0: aoqi@0: default: aoqi@0: fatal_unexpected_iid(id); aoqi@0: break; aoqi@0: } aoqi@0: set_result(_gvn.transform(result)); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: #ifdef _LP64 aoqi@0: #define XTOP ,top() /*additional argument*/ aoqi@0: #else //_LP64 aoqi@0: #define XTOP /*no additional argument*/ aoqi@0: #endif //_LP64 aoqi@0: aoqi@0: //----------------------inline_unsafe_copyMemory------------------------- aoqi@0: // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); aoqi@0: bool LibraryCallKit::inline_unsafe_copyMemory() { aoqi@0: if (callee()->is_static()) return false; // caller must have the capability! aoqi@0: null_check_receiver(); // null-check receiver aoqi@0: if (stopped()) return true; aoqi@0: aoqi@0: C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". aoqi@0: aoqi@0: Node* src_ptr = argument(1); // type: oop aoqi@0: Node* src_off = ConvL2X(argument(2)); // type: long aoqi@0: Node* dst_ptr = argument(4); // type: oop aoqi@0: Node* dst_off = ConvL2X(argument(5)); // type: long aoqi@0: Node* size = ConvL2X(argument(7)); // type: long aoqi@0: aoqi@0: assert(Unsafe_field_offset_to_byte_offset(11) == 11, aoqi@0: "fieldOffset must be byte-scaled"); aoqi@0: aoqi@0: Node* src = make_unsafe_address(src_ptr, src_off); aoqi@0: Node* dst = make_unsafe_address(dst_ptr, dst_off); aoqi@0: aoqi@0: // Conservatively insert a memory barrier on all memory slices. aoqi@0: // Do not let writes of the copy source or destination float below the copy. aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: aoqi@0: // Call it. Note that the length argument is not scaled. aoqi@0: make_runtime_call(RC_LEAF|RC_NO_FP, aoqi@0: OptoRuntime::fast_arraycopy_Type(), aoqi@0: StubRoutines::unsafe_arraycopy(), aoqi@0: "unsafe_arraycopy", aoqi@0: TypeRawPtr::BOTTOM, aoqi@0: src, dst, size XTOP); aoqi@0: aoqi@0: // Do not let reads of the copy destination float above the copy. aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------clone_coping----------------------------------- aoqi@0: // Helper function for inline_native_clone. aoqi@0: void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { aoqi@0: assert(obj_size != NULL, ""); aoqi@0: Node* raw_obj = alloc_obj->in(1); aoqi@0: assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); aoqi@0: aoqi@0: AllocateNode* alloc = NULL; aoqi@0: if (ReduceBulkZeroing) { aoqi@0: // We will be completely responsible for initializing this object - aoqi@0: // mark Initialize node as complete. aoqi@0: alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); aoqi@0: // The object was just allocated - there should be no any stores! aoqi@0: guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); aoqi@0: // Mark as complete_with_arraycopy so that on AllocateNode aoqi@0: // expansion, we know this AllocateNode is initialized by an array aoqi@0: // copy and a StoreStore barrier exists after the array copy. aoqi@0: alloc->initialization()->set_complete_with_arraycopy(); aoqi@0: } aoqi@0: aoqi@0: // Copy the fastest available way. aoqi@0: // TODO: generate fields copies for small objects instead. aoqi@0: Node* src = obj; aoqi@0: Node* dest = alloc_obj; aoqi@0: Node* size = _gvn.transform(obj_size); aoqi@0: aoqi@0: // Exclude the header but include array length to copy by 8 bytes words. aoqi@0: // Can't use base_offset_in_bytes(bt) since basic type is unknown. aoqi@0: int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : aoqi@0: instanceOopDesc::base_offset_in_bytes(); aoqi@0: // base_off: aoqi@0: // 8 - 32-bit VM aoqi@0: // 12 - 64-bit VM, compressed klass aoqi@0: // 16 - 64-bit VM, normal klass aoqi@0: if (base_off % BytesPerLong != 0) { aoqi@0: assert(UseCompressedClassPointers, ""); aoqi@0: if (is_array) { aoqi@0: // Exclude length to copy by 8 bytes words. aoqi@0: base_off += sizeof(int); aoqi@0: } else { aoqi@0: // Include klass to copy by 8 bytes words. aoqi@0: base_off = instanceOopDesc::klass_offset_in_bytes(); aoqi@0: } aoqi@0: assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); aoqi@0: } aoqi@0: src = basic_plus_adr(src, base_off); aoqi@0: dest = basic_plus_adr(dest, base_off); aoqi@0: aoqi@0: // Compute the length also, if needed: aoqi@0: Node* countx = size; aoqi@0: countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off))); aoqi@0: countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) )); aoqi@0: aoqi@0: const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; aoqi@0: bool disjoint_bases = true; aoqi@0: generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, aoqi@0: src, NULL, dest, NULL, countx, aoqi@0: /*dest_uninitialized*/true); aoqi@0: aoqi@0: // If necessary, emit some card marks afterwards. (Non-arrays only.) aoqi@0: if (card_mark) { aoqi@0: assert(!is_array, ""); aoqi@0: // Put in store barrier for any and all oops we are sticking aoqi@0: // into this object. (We could avoid this if we could prove aoqi@0: // that the object type contains no oop fields at all.) aoqi@0: Node* no_particular_value = NULL; aoqi@0: Node* no_particular_field = NULL; aoqi@0: int raw_adr_idx = Compile::AliasIdxRaw; aoqi@0: post_barrier(control(), aoqi@0: memory(raw_adr_type), aoqi@0: alloc_obj, aoqi@0: no_particular_field, aoqi@0: raw_adr_idx, aoqi@0: no_particular_value, aoqi@0: T_OBJECT, aoqi@0: false); aoqi@0: } aoqi@0: aoqi@0: // Do not let reads from the cloned object float above the arraycopy. aoqi@0: if (alloc != NULL) { aoqi@0: // Do not let stores that initialize this object be reordered with aoqi@0: // a subsequent store that would make this object accessible by aoqi@0: // other threads. aoqi@0: // Record what AllocateNode this StoreStore protects so that aoqi@0: // escape analysis can go from the MemBarStoreStoreNode to the aoqi@0: // AllocateNode and eliminate the MemBarStoreStoreNode if possible aoqi@0: // based on the escape status of the AllocateNode. aoqi@0: insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); aoqi@0: } else { aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: //------------------------inline_native_clone---------------------------- aoqi@0: // protected native Object java.lang.Object.clone(); aoqi@0: // aoqi@0: // Here are the simple edge cases: aoqi@0: // null receiver => normal trap aoqi@0: // virtual and clone was overridden => slow path to out-of-line clone aoqi@0: // not cloneable or finalizer => slow path to out-of-line Object.clone aoqi@0: // aoqi@0: // The general case has two steps, allocation and copying. aoqi@0: // Allocation has two cases, and uses GraphKit::new_instance or new_array. aoqi@0: // aoqi@0: // Copying also has two cases, oop arrays and everything else. aoqi@0: // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). aoqi@0: // Everything else uses the tight inline loop supplied by CopyArrayNode. aoqi@0: // aoqi@0: // These steps fold up nicely if and when the cloned object's klass aoqi@0: // can be sharply typed as an object array, a type array, or an instance. aoqi@0: // aoqi@0: bool LibraryCallKit::inline_native_clone(bool is_virtual) { aoqi@0: PhiNode* result_val; aoqi@0: aoqi@0: // Set the reexecute bit for the interpreter to reexecute aoqi@0: // the bytecode that invokes Object.clone if deoptimization happens. aoqi@0: { PreserveReexecuteState preexecs(this); aoqi@0: jvms()->set_should_reexecute(true); aoqi@0: aoqi@0: Node* obj = null_check_receiver(); aoqi@0: if (stopped()) return true; aoqi@0: aoqi@0: Node* obj_klass = load_object_klass(obj); aoqi@0: const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); aoqi@0: const TypeOopPtr* toop = ((tklass != NULL) aoqi@0: ? tklass->as_instance_type() aoqi@0: : TypeInstPtr::NOTNULL); aoqi@0: aoqi@0: // Conservatively insert a memory barrier on all memory slices. aoqi@0: // Do not let writes into the original float below the clone. aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: aoqi@0: // paths into result_reg: aoqi@0: enum { aoqi@0: _slow_path = 1, // out-of-line call to clone method (virtual or not) aoqi@0: _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy aoqi@0: _array_path, // plain array allocation, plus arrayof_long_arraycopy aoqi@0: _instance_path, // plain instance allocation, plus arrayof_long_arraycopy aoqi@0: PATH_LIMIT aoqi@0: }; aoqi@0: RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); aoqi@0: result_val = new(C) PhiNode(result_reg, aoqi@0: TypeInstPtr::NOTNULL); aoqi@0: PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO); aoqi@0: PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, aoqi@0: TypePtr::BOTTOM); aoqi@0: record_for_igvn(result_reg); aoqi@0: aoqi@0: const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; aoqi@0: int raw_adr_idx = Compile::AliasIdxRaw; aoqi@0: aoqi@0: Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); aoqi@0: if (array_ctl != NULL) { aoqi@0: // It's an array. aoqi@0: PreserveJVMState pjvms(this); aoqi@0: set_control(array_ctl); aoqi@0: Node* obj_length = load_array_length(obj); aoqi@0: Node* obj_size = NULL; aoqi@0: Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push aoqi@0: aoqi@0: if (!use_ReduceInitialCardMarks()) { aoqi@0: // If it is an oop array, it requires very special treatment, aoqi@0: // because card marking is required on each card of the array. aoqi@0: Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); aoqi@0: if (is_obja != NULL) { aoqi@0: PreserveJVMState pjvms2(this); aoqi@0: set_control(is_obja); aoqi@0: // Generate a direct call to the right arraycopy function(s). aoqi@0: bool disjoint_bases = true; aoqi@0: bool length_never_negative = true; aoqi@0: generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, aoqi@0: obj, intcon(0), alloc_obj, intcon(0), aoqi@0: obj_length, aoqi@0: disjoint_bases, length_never_negative); aoqi@0: result_reg->init_req(_objArray_path, control()); aoqi@0: result_val->init_req(_objArray_path, alloc_obj); aoqi@0: result_i_o ->set_req(_objArray_path, i_o()); aoqi@0: result_mem ->set_req(_objArray_path, reset_memory()); aoqi@0: } aoqi@0: } aoqi@0: // Otherwise, there are no card marks to worry about. aoqi@0: // (We can dispense with card marks if we know the allocation aoqi@0: // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks aoqi@0: // causes the non-eden paths to take compensating steps to aoqi@0: // simulate a fresh allocation, so that no further aoqi@0: // card marks are required in compiled code to initialize aoqi@0: // the object.) aoqi@0: aoqi@0: if (!stopped()) { aoqi@0: copy_to_clone(obj, alloc_obj, obj_size, true, false); aoqi@0: aoqi@0: // Present the results of the copy. aoqi@0: result_reg->init_req(_array_path, control()); aoqi@0: result_val->init_req(_array_path, alloc_obj); aoqi@0: result_i_o ->set_req(_array_path, i_o()); aoqi@0: result_mem ->set_req(_array_path, reset_memory()); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // We only go to the instance fast case code if we pass a number of guards. aoqi@0: // The paths which do not pass are accumulated in the slow_region. aoqi@0: RegionNode* slow_region = new (C) RegionNode(1); aoqi@0: record_for_igvn(slow_region); aoqi@0: if (!stopped()) { aoqi@0: // It's an instance (we did array above). Make the slow-path tests. aoqi@0: // If this is a virtual call, we generate a funny guard. We grab aoqi@0: // the vtable entry corresponding to clone() from the target object. aoqi@0: // If the target method which we are calling happens to be the aoqi@0: // Object clone() method, we pass the guard. We do not need this aoqi@0: // guard for non-virtual calls; the caller is known to be the native aoqi@0: // Object clone(). aoqi@0: if (is_virtual) { aoqi@0: generate_virtual_guard(obj_klass, slow_region); aoqi@0: } aoqi@0: aoqi@0: // The object must be cloneable and must not have a finalizer. aoqi@0: // Both of these conditions may be checked in a single test. aoqi@0: // We could optimize the cloneable test further, but we don't care. aoqi@0: generate_access_flags_guard(obj_klass, aoqi@0: // Test both conditions: aoqi@0: JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, aoqi@0: // Must be cloneable but not finalizer: aoqi@0: JVM_ACC_IS_CLONEABLE, aoqi@0: slow_region); aoqi@0: } aoqi@0: aoqi@0: if (!stopped()) { aoqi@0: // It's an instance, and it passed the slow-path tests. aoqi@0: PreserveJVMState pjvms(this); aoqi@0: Node* obj_size = NULL; aoqi@0: // Need to deoptimize on exception from allocation since Object.clone intrinsic aoqi@0: // is reexecuted if deoptimization occurs and there could be problems when merging aoqi@0: // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). aoqi@0: Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); aoqi@0: aoqi@0: copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); aoqi@0: aoqi@0: // Present the results of the slow call. aoqi@0: result_reg->init_req(_instance_path, control()); aoqi@0: result_val->init_req(_instance_path, alloc_obj); aoqi@0: result_i_o ->set_req(_instance_path, i_o()); aoqi@0: result_mem ->set_req(_instance_path, reset_memory()); aoqi@0: } aoqi@0: aoqi@0: // Generate code for the slow case. We make a call to clone(). aoqi@0: set_control(_gvn.transform(slow_region)); aoqi@0: if (!stopped()) { aoqi@0: PreserveJVMState pjvms(this); aoqi@0: CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); aoqi@0: Node* slow_result = set_results_for_java_call(slow_call); aoqi@0: // this->control() comes from set_results_for_java_call aoqi@0: result_reg->init_req(_slow_path, control()); aoqi@0: result_val->init_req(_slow_path, slow_result); aoqi@0: result_i_o ->set_req(_slow_path, i_o()); aoqi@0: result_mem ->set_req(_slow_path, reset_memory()); aoqi@0: } aoqi@0: aoqi@0: // Return the combined state. aoqi@0: set_control( _gvn.transform(result_reg)); aoqi@0: set_i_o( _gvn.transform(result_i_o)); aoqi@0: set_all_memory( _gvn.transform(result_mem)); aoqi@0: } // original reexecute is set back here aoqi@0: aoqi@0: set_result(_gvn.transform(result_val)); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------------basictype2arraycopy---------------------------- aoqi@0: address LibraryCallKit::basictype2arraycopy(BasicType t, aoqi@0: Node* src_offset, aoqi@0: Node* dest_offset, aoqi@0: bool disjoint_bases, aoqi@0: const char* &name, aoqi@0: bool dest_uninitialized) { aoqi@0: const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; aoqi@0: const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; aoqi@0: aoqi@0: bool aligned = false; aoqi@0: bool disjoint = disjoint_bases; aoqi@0: aoqi@0: // if the offsets are the same, we can treat the memory regions as aoqi@0: // disjoint, because either the memory regions are in different arrays, aoqi@0: // or they are identical (which we can treat as disjoint.) We can also aoqi@0: // treat a copy with a destination index less that the source index aoqi@0: // as disjoint since a low->high copy will work correctly in this case. aoqi@0: if (src_offset_inttype != NULL && src_offset_inttype->is_con() && aoqi@0: dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { aoqi@0: // both indices are constants aoqi@0: int s_offs = src_offset_inttype->get_con(); aoqi@0: int d_offs = dest_offset_inttype->get_con(); aoqi@0: int element_size = type2aelembytes(t); aoqi@0: aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && aoqi@0: ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); aoqi@0: if (s_offs >= d_offs) disjoint = true; aoqi@0: } else if (src_offset == dest_offset && src_offset != NULL) { aoqi@0: // This can occur if the offsets are identical non-constants. aoqi@0: disjoint = true; aoqi@0: } aoqi@0: aoqi@0: return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); aoqi@0: } aoqi@0: aoqi@0: aoqi@0: //------------------------------inline_arraycopy----------------------- aoqi@0: // public static native void java.lang.System.arraycopy(Object src, int srcPos, aoqi@0: // Object dest, int destPos, aoqi@0: // int length); aoqi@0: bool LibraryCallKit::inline_arraycopy() { aoqi@0: // Get the arguments. aoqi@0: Node* src = argument(0); // type: oop aoqi@0: Node* src_offset = argument(1); // type: int aoqi@0: Node* dest = argument(2); // type: oop aoqi@0: Node* dest_offset = argument(3); // type: int aoqi@0: Node* length = argument(4); // type: int aoqi@0: aoqi@0: // Compile time checks. If any of these checks cannot be verified at compile time, aoqi@0: // we do not make a fast path for this call. Instead, we let the call remain as it aoqi@0: // is. The checks we choose to mandate at compile time are: aoqi@0: // aoqi@0: // (1) src and dest are arrays. aoqi@0: const Type* src_type = src->Value(&_gvn); aoqi@0: const Type* dest_type = dest->Value(&_gvn); aoqi@0: const TypeAryPtr* top_src = src_type->isa_aryptr(); aoqi@0: const TypeAryPtr* top_dest = dest_type->isa_aryptr(); aoqi@0: aoqi@0: // Do we have the type of src? aoqi@0: bool has_src = (top_src != NULL && top_src->klass() != NULL); aoqi@0: // Do we have the type of dest? aoqi@0: bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); aoqi@0: // Is the type for src from speculation? aoqi@0: bool src_spec = false; aoqi@0: // Is the type for dest from speculation? aoqi@0: bool dest_spec = false; aoqi@0: aoqi@0: if (!has_src || !has_dest) { aoqi@0: // We don't have sufficient type information, let's see if aoqi@0: // speculative types can help. We need to have types for both src aoqi@0: // and dest so that it pays off. aoqi@0: aoqi@0: // Do we already have or could we have type information for src aoqi@0: bool could_have_src = has_src; aoqi@0: // Do we already have or could we have type information for dest aoqi@0: bool could_have_dest = has_dest; aoqi@0: aoqi@0: ciKlass* src_k = NULL; aoqi@0: if (!has_src) { aoqi@0: src_k = src_type->speculative_type(); aoqi@0: if (src_k != NULL && src_k->is_array_klass()) { aoqi@0: could_have_src = true; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: ciKlass* dest_k = NULL; aoqi@0: if (!has_dest) { aoqi@0: dest_k = dest_type->speculative_type(); aoqi@0: if (dest_k != NULL && dest_k->is_array_klass()) { aoqi@0: could_have_dest = true; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: if (could_have_src && could_have_dest) { aoqi@0: // This is going to pay off so emit the required guards aoqi@0: if (!has_src) { aoqi@0: src = maybe_cast_profiled_obj(src, src_k); aoqi@0: src_type = _gvn.type(src); aoqi@0: top_src = src_type->isa_aryptr(); aoqi@0: has_src = (top_src != NULL && top_src->klass() != NULL); aoqi@0: src_spec = true; aoqi@0: } aoqi@0: if (!has_dest) { aoqi@0: dest = maybe_cast_profiled_obj(dest, dest_k); aoqi@0: dest_type = _gvn.type(dest); aoqi@0: top_dest = dest_type->isa_aryptr(); aoqi@0: has_dest = (top_dest != NULL && top_dest->klass() != NULL); aoqi@0: dest_spec = true; aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: if (!has_src || !has_dest) { aoqi@0: // Conservatively insert a memory barrier on all memory slices. aoqi@0: // Do not let writes into the source float below the arraycopy. aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: aoqi@0: // Call StubRoutines::generic_arraycopy stub. aoqi@0: generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, aoqi@0: src, src_offset, dest, dest_offset, length); aoqi@0: aoqi@0: // Do not let reads from the destination float above the arraycopy. aoqi@0: // Since we cannot type the arrays, we don't know which slices aoqi@0: // might be affected. We could restrict this barrier only to those aoqi@0: // memory slices which pertain to array elements--but don't bother. aoqi@0: if (!InsertMemBarAfterArraycopy) aoqi@0: // (If InsertMemBarAfterArraycopy, there is already one in place.) aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: // (2) src and dest arrays must have elements of the same BasicType aoqi@0: // Figure out the size and type of the elements we will be copying. aoqi@0: BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); aoqi@0: BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); aoqi@0: if (src_elem == T_ARRAY) src_elem = T_OBJECT; aoqi@0: if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; aoqi@0: aoqi@0: if (src_elem != dest_elem || dest_elem == T_VOID) { aoqi@0: // The component types are not the same or are not recognized. Punt. aoqi@0: // (But, avoid the native method wrapper to JVM_ArrayCopy.) aoqi@0: generate_slow_arraycopy(TypePtr::BOTTOM, aoqi@0: src, src_offset, dest, dest_offset, length, aoqi@0: /*dest_uninitialized*/false); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: if (src_elem == T_OBJECT) { aoqi@0: // If both arrays are object arrays then having the exact types aoqi@0: // for both will remove the need for a subtype check at runtime aoqi@0: // before the call and may make it possible to pick a faster copy aoqi@0: // routine (without a subtype check on every element) aoqi@0: // Do we have the exact type of src? aoqi@0: bool could_have_src = src_spec; aoqi@0: // Do we have the exact type of dest? aoqi@0: bool could_have_dest = dest_spec; aoqi@0: ciKlass* src_k = top_src->klass(); aoqi@0: ciKlass* dest_k = top_dest->klass(); aoqi@0: if (!src_spec) { aoqi@0: src_k = src_type->speculative_type(); aoqi@0: if (src_k != NULL && src_k->is_array_klass()) { aoqi@0: could_have_src = true; aoqi@0: } aoqi@0: } aoqi@0: if (!dest_spec) { aoqi@0: dest_k = dest_type->speculative_type(); aoqi@0: if (dest_k != NULL && dest_k->is_array_klass()) { aoqi@0: could_have_dest = true; aoqi@0: } aoqi@0: } aoqi@0: if (could_have_src && could_have_dest) { aoqi@0: // If we can have both exact types, emit the missing guards aoqi@0: if (could_have_src && !src_spec) { aoqi@0: src = maybe_cast_profiled_obj(src, src_k); aoqi@0: } aoqi@0: if (could_have_dest && !dest_spec) { aoqi@0: dest = maybe_cast_profiled_obj(dest, dest_k); aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: //--------------------------------------------------------------------------- aoqi@0: // We will make a fast path for this call to arraycopy. aoqi@0: aoqi@0: // We have the following tests left to perform: aoqi@0: // aoqi@0: // (3) src and dest must not be null. aoqi@0: // (4) src_offset must not be negative. aoqi@0: // (5) dest_offset must not be negative. aoqi@0: // (6) length must not be negative. aoqi@0: // (7) src_offset + length must not exceed length of src. aoqi@0: // (8) dest_offset + length must not exceed length of dest. aoqi@0: // (9) each element of an oop array must be assignable aoqi@0: aoqi@0: RegionNode* slow_region = new (C) RegionNode(1); aoqi@0: record_for_igvn(slow_region); aoqi@0: aoqi@0: // (3) operands must not be null aoqi@0: // We currently perform our null checks with the null_check routine. aoqi@0: // This means that the null exceptions will be reported in the caller aoqi@0: // rather than (correctly) reported inside of the native arraycopy call. aoqi@0: // This should be corrected, given time. We do our null check with the aoqi@0: // stack pointer restored. aoqi@0: src = null_check(src, T_ARRAY); aoqi@0: dest = null_check(dest, T_ARRAY); aoqi@0: aoqi@0: // (4) src_offset must not be negative. aoqi@0: generate_negative_guard(src_offset, slow_region); aoqi@0: aoqi@0: // (5) dest_offset must not be negative. aoqi@0: generate_negative_guard(dest_offset, slow_region); aoqi@0: aoqi@0: // (6) length must not be negative (moved to generate_arraycopy()). aoqi@0: // generate_negative_guard(length, slow_region); aoqi@0: aoqi@0: // (7) src_offset + length must not exceed length of src. aoqi@0: generate_limit_guard(src_offset, length, aoqi@0: load_array_length(src), aoqi@0: slow_region); aoqi@0: aoqi@0: // (8) dest_offset + length must not exceed length of dest. aoqi@0: generate_limit_guard(dest_offset, length, aoqi@0: load_array_length(dest), aoqi@0: slow_region); aoqi@0: aoqi@0: // (9) each element of an oop array must be assignable aoqi@0: // The generate_arraycopy subroutine checks this. aoqi@0: aoqi@0: // This is where the memory effects are placed: aoqi@0: const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); aoqi@0: generate_arraycopy(adr_type, dest_elem, aoqi@0: src, src_offset, dest, dest_offset, length, aoqi@0: false, false, slow_region); aoqi@0: aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //-----------------------------generate_arraycopy---------------------- aoqi@0: // Generate an optimized call to arraycopy. aoqi@0: // Caller must guard against non-arrays. aoqi@0: // Caller must determine a common array basic-type for both arrays. aoqi@0: // Caller must validate offsets against array bounds. aoqi@0: // The slow_region has already collected guard failure paths aoqi@0: // (such as out of bounds length or non-conformable array types). aoqi@0: // The generated code has this shape, in general: aoqi@0: // aoqi@0: // if (length == 0) return // via zero_path aoqi@0: // slowval = -1 aoqi@0: // if (types unknown) { aoqi@0: // slowval = call generic copy loop aoqi@0: // if (slowval == 0) return // via checked_path aoqi@0: // } else if (indexes in bounds) { aoqi@0: // if ((is object array) && !(array type check)) { aoqi@0: // slowval = call checked copy loop aoqi@0: // if (slowval == 0) return // via checked_path aoqi@0: // } else { aoqi@0: // call bulk copy loop aoqi@0: // return // via fast_path aoqi@0: // } aoqi@0: // } aoqi@0: // // adjust params for remaining work: aoqi@0: // if (slowval != -1) { aoqi@0: // n = -1^slowval; src_offset += n; dest_offset += n; length -= n aoqi@0: // } aoqi@0: // slow_region: aoqi@0: // call slow arraycopy(src, src_offset, dest, dest_offset, length) aoqi@0: // return // via slow_call_path aoqi@0: // aoqi@0: // This routine is used from several intrinsics: System.arraycopy, aoqi@0: // Object.clone (the array subcase), and Arrays.copyOf[Range]. aoqi@0: // aoqi@0: void aoqi@0: LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, aoqi@0: BasicType basic_elem_type, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* copy_length, aoqi@0: bool disjoint_bases, aoqi@0: bool length_never_negative, aoqi@0: RegionNode* slow_region) { aoqi@0: aoqi@0: if (slow_region == NULL) { aoqi@0: slow_region = new(C) RegionNode(1); aoqi@0: record_for_igvn(slow_region); aoqi@0: } aoqi@0: aoqi@0: Node* original_dest = dest; aoqi@0: AllocateArrayNode* alloc = NULL; // used for zeroing, if needed aoqi@0: bool dest_uninitialized = false; aoqi@0: aoqi@0: // See if this is the initialization of a newly-allocated array. aoqi@0: // If so, we will take responsibility here for initializing it to zero. aoqi@0: // (Note: Because tightly_coupled_allocation performs checks on the aoqi@0: // out-edges of the dest, we need to avoid making derived pointers aoqi@0: // from it until we have checked its uses.) aoqi@0: if (ReduceBulkZeroing aoqi@0: && !ZeroTLAB // pointless if already zeroed aoqi@0: && basic_elem_type != T_CONFLICT // avoid corner case aoqi@0: && !src->eqv_uncast(dest) aoqi@0: && ((alloc = tightly_coupled_allocation(dest, slow_region)) aoqi@0: != NULL) aoqi@0: && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 aoqi@0: && alloc->maybe_set_complete(&_gvn)) { aoqi@0: // "You break it, you buy it." aoqi@0: InitializeNode* init = alloc->initialization(); aoqi@0: assert(init->is_complete(), "we just did this"); aoqi@0: init->set_complete_with_arraycopy(); aoqi@0: assert(dest->is_CheckCastPP(), "sanity"); aoqi@0: assert(dest->in(0)->in(0) == init, "dest pinned"); aoqi@0: adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory aoqi@0: // From this point on, every exit path is responsible for aoqi@0: // initializing any non-copied parts of the object to zero. aoqi@0: // Also, if this flag is set we make sure that arraycopy interacts properly aoqi@0: // with G1, eliding pre-barriers. See CR 6627983. aoqi@0: dest_uninitialized = true; aoqi@0: } else { aoqi@0: // No zeroing elimination here. aoqi@0: alloc = NULL; aoqi@0: //original_dest = dest; aoqi@0: //dest_uninitialized = false; aoqi@0: } aoqi@0: aoqi@0: // Results are placed here: aoqi@0: enum { fast_path = 1, // normal void-returning assembly stub aoqi@0: checked_path = 2, // special assembly stub with cleanup aoqi@0: slow_call_path = 3, // something went wrong; call the VM aoqi@0: zero_path = 4, // bypass when length of copy is zero aoqi@0: bcopy_path = 5, // copy primitive array by 64-bit blocks aoqi@0: PATH_LIMIT = 6 aoqi@0: }; aoqi@0: RegionNode* result_region = new(C) RegionNode(PATH_LIMIT); aoqi@0: PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO); aoqi@0: PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type); aoqi@0: record_for_igvn(result_region); aoqi@0: _gvn.set_type_bottom(result_i_o); aoqi@0: _gvn.set_type_bottom(result_memory); aoqi@0: assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); aoqi@0: aoqi@0: // The slow_control path: aoqi@0: Node* slow_control; aoqi@0: Node* slow_i_o = i_o(); aoqi@0: Node* slow_mem = memory(adr_type); aoqi@0: debug_only(slow_control = (Node*) badAddress); aoqi@0: aoqi@0: // Checked control path: aoqi@0: Node* checked_control = top(); aoqi@0: Node* checked_mem = NULL; aoqi@0: Node* checked_i_o = NULL; aoqi@0: Node* checked_value = NULL; aoqi@0: aoqi@0: if (basic_elem_type == T_CONFLICT) { aoqi@0: assert(!dest_uninitialized, ""); aoqi@0: Node* cv = generate_generic_arraycopy(adr_type, aoqi@0: src, src_offset, dest, dest_offset, aoqi@0: copy_length, dest_uninitialized); aoqi@0: if (cv == NULL) cv = intcon(-1); // failure (no stub available) aoqi@0: checked_control = control(); aoqi@0: checked_i_o = i_o(); aoqi@0: checked_mem = memory(adr_type); aoqi@0: checked_value = cv; aoqi@0: set_control(top()); // no fast path aoqi@0: } aoqi@0: aoqi@0: Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); aoqi@0: if (not_pos != NULL) { aoqi@0: PreserveJVMState pjvms(this); aoqi@0: set_control(not_pos); aoqi@0: aoqi@0: // (6) length must not be negative. aoqi@0: if (!length_never_negative) { aoqi@0: generate_negative_guard(copy_length, slow_region); aoqi@0: } aoqi@0: aoqi@0: // copy_length is 0. aoqi@0: if (!stopped() && dest_uninitialized) { aoqi@0: Node* dest_length = alloc->in(AllocateNode::ALength); aoqi@0: if (copy_length->eqv_uncast(dest_length) aoqi@0: || _gvn.find_int_con(dest_length, 1) <= 0) { aoqi@0: // There is no zeroing to do. No need for a secondary raw memory barrier. aoqi@0: } else { aoqi@0: // Clear the whole thing since there are no source elements to copy. aoqi@0: generate_clear_array(adr_type, dest, basic_elem_type, aoqi@0: intcon(0), NULL, aoqi@0: alloc->in(AllocateNode::AllocSize)); aoqi@0: // Use a secondary InitializeNode as raw memory barrier. aoqi@0: // Currently it is needed only on this path since other aoqi@0: // paths have stub or runtime calls as raw memory barriers. aoqi@0: InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, aoqi@0: Compile::AliasIdxRaw, aoqi@0: top())->as_Initialize(); aoqi@0: init->set_complete(&_gvn); // (there is no corresponding AllocateNode) aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // Present the results of the fast call. aoqi@0: result_region->init_req(zero_path, control()); aoqi@0: result_i_o ->init_req(zero_path, i_o()); aoqi@0: result_memory->init_req(zero_path, memory(adr_type)); aoqi@0: } aoqi@0: aoqi@0: if (!stopped() && dest_uninitialized) { aoqi@0: // We have to initialize the *uncopied* part of the array to zero. aoqi@0: // The copy destination is the slice dest[off..off+len]. The other slices aoqi@0: // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. aoqi@0: Node* dest_size = alloc->in(AllocateNode::AllocSize); aoqi@0: Node* dest_length = alloc->in(AllocateNode::ALength); aoqi@0: Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset, aoqi@0: copy_length)); aoqi@0: aoqi@0: // If there is a head section that needs zeroing, do it now. aoqi@0: if (find_int_con(dest_offset, -1) != 0) { aoqi@0: generate_clear_array(adr_type, dest, basic_elem_type, aoqi@0: intcon(0), dest_offset, aoqi@0: NULL); aoqi@0: } aoqi@0: aoqi@0: // Next, perform a dynamic check on the tail length. aoqi@0: // It is often zero, and we can win big if we prove this. aoqi@0: // There are two wins: Avoid generating the ClearArray aoqi@0: // with its attendant messy index arithmetic, and upgrade aoqi@0: // the copy to a more hardware-friendly word size of 64 bits. aoqi@0: Node* tail_ctl = NULL; aoqi@0: if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { aoqi@0: Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length)); aoqi@0: Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt)); aoqi@0: tail_ctl = generate_slow_guard(bol_lt, NULL); aoqi@0: assert(tail_ctl != NULL || !stopped(), "must be an outcome"); aoqi@0: } aoqi@0: aoqi@0: // At this point, let's assume there is no tail. aoqi@0: if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { aoqi@0: // There is no tail. Try an upgrade to a 64-bit copy. aoqi@0: bool didit = false; aoqi@0: { PreserveJVMState pjvms(this); aoqi@0: didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, aoqi@0: src, src_offset, dest, dest_offset, aoqi@0: dest_size, dest_uninitialized); aoqi@0: if (didit) { aoqi@0: // Present the results of the block-copying fast call. aoqi@0: result_region->init_req(bcopy_path, control()); aoqi@0: result_i_o ->init_req(bcopy_path, i_o()); aoqi@0: result_memory->init_req(bcopy_path, memory(adr_type)); aoqi@0: } aoqi@0: } aoqi@0: if (didit) aoqi@0: set_control(top()); // no regular fast path aoqi@0: } aoqi@0: aoqi@0: // Clear the tail, if any. aoqi@0: if (tail_ctl != NULL) { aoqi@0: Node* notail_ctl = stopped() ? NULL : control(); aoqi@0: set_control(tail_ctl); aoqi@0: if (notail_ctl == NULL) { aoqi@0: generate_clear_array(adr_type, dest, basic_elem_type, aoqi@0: dest_tail, NULL, aoqi@0: dest_size); aoqi@0: } else { aoqi@0: // Make a local merge. aoqi@0: Node* done_ctl = new(C) RegionNode(3); aoqi@0: Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type); aoqi@0: done_ctl->init_req(1, notail_ctl); aoqi@0: done_mem->init_req(1, memory(adr_type)); aoqi@0: generate_clear_array(adr_type, dest, basic_elem_type, aoqi@0: dest_tail, NULL, aoqi@0: dest_size); aoqi@0: done_ctl->init_req(2, control()); aoqi@0: done_mem->init_req(2, memory(adr_type)); aoqi@0: set_control( _gvn.transform(done_ctl)); aoqi@0: set_memory( _gvn.transform(done_mem), adr_type ); aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: BasicType copy_type = basic_elem_type; aoqi@0: assert(basic_elem_type != T_ARRAY, "caller must fix this"); aoqi@0: if (!stopped() && copy_type == T_OBJECT) { aoqi@0: // If src and dest have compatible element types, we can copy bits. aoqi@0: // Types S[] and D[] are compatible if D is a supertype of S. aoqi@0: // aoqi@0: // If they are not, we will use checked_oop_disjoint_arraycopy, aoqi@0: // which performs a fast optimistic per-oop check, and backs off aoqi@0: // further to JVM_ArrayCopy on the first per-oop check that fails. aoqi@0: // (Actually, we don't move raw bits only; the GC requires card marks.) aoqi@0: aoqi@0: // Get the Klass* for both src and dest aoqi@0: Node* src_klass = load_object_klass(src); aoqi@0: Node* dest_klass = load_object_klass(dest); aoqi@0: aoqi@0: // Generate the subtype check. aoqi@0: // This might fold up statically, or then again it might not. aoqi@0: // aoqi@0: // Non-static example: Copying List.elements to a new String[]. aoqi@0: // The backing store for a List is always an Object[], aoqi@0: // but its elements are always type String, if the generic types aoqi@0: // are correct at the source level. aoqi@0: // aoqi@0: // Test S[] against D[], not S against D, because (probably) aoqi@0: // the secondary supertype cache is less busy for S[] than S. aoqi@0: // This usually only matters when D is an interface. aoqi@0: Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); aoqi@0: // Plug failing path into checked_oop_disjoint_arraycopy aoqi@0: if (not_subtype_ctrl != top()) { aoqi@0: PreserveJVMState pjvms(this); aoqi@0: set_control(not_subtype_ctrl); aoqi@0: // (At this point we can assume disjoint_bases, since types differ.) aoqi@0: int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); aoqi@0: Node* p1 = basic_plus_adr(dest_klass, ek_offset); aoqi@0: Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); aoqi@0: Node* dest_elem_klass = _gvn.transform(n1); aoqi@0: Node* cv = generate_checkcast_arraycopy(adr_type, aoqi@0: dest_elem_klass, aoqi@0: src, src_offset, dest, dest_offset, aoqi@0: ConvI2X(copy_length), dest_uninitialized); aoqi@0: if (cv == NULL) cv = intcon(-1); // failure (no stub available) aoqi@0: checked_control = control(); aoqi@0: checked_i_o = i_o(); aoqi@0: checked_mem = memory(adr_type); aoqi@0: checked_value = cv; aoqi@0: } aoqi@0: // At this point we know we do not need type checks on oop stores. aoqi@0: aoqi@0: // Let's see if we need card marks: aoqi@0: if (alloc != NULL && use_ReduceInitialCardMarks()) { aoqi@0: // If we do not need card marks, copy using the jint or jlong stub. aoqi@0: copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); aoqi@0: assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), aoqi@0: "sizes agree"); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: if (!stopped()) { aoqi@0: // Generate the fast path, if possible. aoqi@0: PreserveJVMState pjvms(this); aoqi@0: generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, aoqi@0: src, src_offset, dest, dest_offset, aoqi@0: ConvI2X(copy_length), dest_uninitialized); aoqi@0: aoqi@0: // Present the results of the fast call. aoqi@0: result_region->init_req(fast_path, control()); aoqi@0: result_i_o ->init_req(fast_path, i_o()); aoqi@0: result_memory->init_req(fast_path, memory(adr_type)); aoqi@0: } aoqi@0: aoqi@0: // Here are all the slow paths up to this point, in one bundle: aoqi@0: slow_control = top(); aoqi@0: if (slow_region != NULL) aoqi@0: slow_control = _gvn.transform(slow_region); aoqi@0: DEBUG_ONLY(slow_region = (RegionNode*)badAddress); aoqi@0: aoqi@0: set_control(checked_control); aoqi@0: if (!stopped()) { aoqi@0: // Clean up after the checked call. aoqi@0: // The returned value is either 0 or -1^K, aoqi@0: // where K = number of partially transferred array elements. aoqi@0: Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0))); aoqi@0: Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); aoqi@0: IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); aoqi@0: aoqi@0: // If it is 0, we are done, so transfer to the end. aoqi@0: Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff)); aoqi@0: result_region->init_req(checked_path, checks_done); aoqi@0: result_i_o ->init_req(checked_path, checked_i_o); aoqi@0: result_memory->init_req(checked_path, checked_mem); aoqi@0: aoqi@0: // If it is not zero, merge into the slow call. aoqi@0: set_control( _gvn.transform(new(C) IfFalseNode(iff) )); aoqi@0: RegionNode* slow_reg2 = new(C) RegionNode(3); aoqi@0: PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO); aoqi@0: PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type); aoqi@0: record_for_igvn(slow_reg2); aoqi@0: slow_reg2 ->init_req(1, slow_control); aoqi@0: slow_i_o2 ->init_req(1, slow_i_o); aoqi@0: slow_mem2 ->init_req(1, slow_mem); aoqi@0: slow_reg2 ->init_req(2, control()); aoqi@0: slow_i_o2 ->init_req(2, checked_i_o); aoqi@0: slow_mem2 ->init_req(2, checked_mem); aoqi@0: aoqi@0: slow_control = _gvn.transform(slow_reg2); aoqi@0: slow_i_o = _gvn.transform(slow_i_o2); aoqi@0: slow_mem = _gvn.transform(slow_mem2); aoqi@0: aoqi@0: if (alloc != NULL) { aoqi@0: // We'll restart from the very beginning, after zeroing the whole thing. aoqi@0: // This can cause double writes, but that's OK since dest is brand new. aoqi@0: // So we ignore the low 31 bits of the value returned from the stub. aoqi@0: } else { aoqi@0: // We must continue the copy exactly where it failed, or else aoqi@0: // another thread might see the wrong number of writes to dest. aoqi@0: Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1))); aoqi@0: Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT); aoqi@0: slow_offset->init_req(1, intcon(0)); aoqi@0: slow_offset->init_req(2, checked_offset); aoqi@0: slow_offset = _gvn.transform(slow_offset); aoqi@0: aoqi@0: // Adjust the arguments by the conditionally incoming offset. aoqi@0: Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset)); aoqi@0: Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset)); aoqi@0: Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset)); aoqi@0: aoqi@0: // Tweak the node variables to adjust the code produced below: aoqi@0: src_offset = src_off_plus; aoqi@0: dest_offset = dest_off_plus; aoqi@0: copy_length = length_minus; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: set_control(slow_control); aoqi@0: if (!stopped()) { aoqi@0: // Generate the slow path, if needed. aoqi@0: PreserveJVMState pjvms(this); // replace_in_map may trash the map aoqi@0: aoqi@0: set_memory(slow_mem, adr_type); aoqi@0: set_i_o(slow_i_o); aoqi@0: aoqi@0: if (dest_uninitialized) { aoqi@0: generate_clear_array(adr_type, dest, basic_elem_type, aoqi@0: intcon(0), NULL, aoqi@0: alloc->in(AllocateNode::AllocSize)); aoqi@0: } aoqi@0: aoqi@0: generate_slow_arraycopy(adr_type, aoqi@0: src, src_offset, dest, dest_offset, aoqi@0: copy_length, /*dest_uninitialized*/false); aoqi@0: aoqi@0: result_region->init_req(slow_call_path, control()); aoqi@0: result_i_o ->init_req(slow_call_path, i_o()); aoqi@0: result_memory->init_req(slow_call_path, memory(adr_type)); aoqi@0: } aoqi@0: aoqi@0: // Remove unused edges. aoqi@0: for (uint i = 1; i < result_region->req(); i++) { aoqi@0: if (result_region->in(i) == NULL) aoqi@0: result_region->init_req(i, top()); aoqi@0: } aoqi@0: aoqi@0: // Finished; return the combined state. aoqi@0: set_control( _gvn.transform(result_region)); aoqi@0: set_i_o( _gvn.transform(result_i_o) ); aoqi@0: set_memory( _gvn.transform(result_memory), adr_type ); aoqi@0: aoqi@0: // The memory edges above are precise in order to model effects around aoqi@0: // array copies accurately to allow value numbering of field loads around aoqi@0: // arraycopy. Such field loads, both before and after, are common in Java aoqi@0: // collections and similar classes involving header/array data structures. aoqi@0: // aoqi@0: // But with low number of register or when some registers are used or killed aoqi@0: // by arraycopy calls it causes registers spilling on stack. See 6544710. aoqi@0: // The next memory barrier is added to avoid it. If the arraycopy can be aoqi@0: // optimized away (which it can, sometimes) then we can manually remove aoqi@0: // the membar also. aoqi@0: // aoqi@0: // Do not let reads from the cloned object float above the arraycopy. aoqi@0: if (alloc != NULL) { aoqi@0: // Do not let stores that initialize this object be reordered with aoqi@0: // a subsequent store that would make this object accessible by aoqi@0: // other threads. aoqi@0: // Record what AllocateNode this StoreStore protects so that aoqi@0: // escape analysis can go from the MemBarStoreStoreNode to the aoqi@0: // AllocateNode and eliminate the MemBarStoreStoreNode if possible aoqi@0: // based on the escape status of the AllocateNode. aoqi@0: insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); aoqi@0: } else if (InsertMemBarAfterArraycopy) aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: } aoqi@0: aoqi@0: aoqi@0: // Helper function which determines if an arraycopy immediately follows aoqi@0: // an allocation, with no intervening tests or other escapes for the object. aoqi@0: AllocateArrayNode* aoqi@0: LibraryCallKit::tightly_coupled_allocation(Node* ptr, aoqi@0: RegionNode* slow_region) { aoqi@0: if (stopped()) return NULL; // no fast path aoqi@0: if (C->AliasLevel() == 0) return NULL; // no MergeMems around aoqi@0: aoqi@0: AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); aoqi@0: if (alloc == NULL) return NULL; aoqi@0: aoqi@0: Node* rawmem = memory(Compile::AliasIdxRaw); aoqi@0: // Is the allocation's memory state untouched? aoqi@0: if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { aoqi@0: // Bail out if there have been raw-memory effects since the allocation. aoqi@0: // (Example: There might have been a call or safepoint.) aoqi@0: return NULL; aoqi@0: } aoqi@0: rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); aoqi@0: if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: // There must be no unexpected observers of this allocation. aoqi@0: for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { aoqi@0: Node* obs = ptr->fast_out(i); aoqi@0: if (obs != this->map()) { aoqi@0: return NULL; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // This arraycopy must unconditionally follow the allocation of the ptr. aoqi@0: Node* alloc_ctl = ptr->in(0); aoqi@0: assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); aoqi@0: aoqi@0: Node* ctl = control(); aoqi@0: while (ctl != alloc_ctl) { aoqi@0: // There may be guards which feed into the slow_region. aoqi@0: // Any other control flow means that we might not get a chance aoqi@0: // to finish initializing the allocated object. aoqi@0: if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { aoqi@0: IfNode* iff = ctl->in(0)->as_If(); aoqi@0: Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); aoqi@0: assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); aoqi@0: if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { aoqi@0: ctl = iff->in(0); // This test feeds the known slow_region. aoqi@0: continue; aoqi@0: } aoqi@0: // One more try: Various low-level checks bottom out in aoqi@0: // uncommon traps. If the debug-info of the trap omits aoqi@0: // any reference to the allocation, as we've already aoqi@0: // observed, then there can be no objection to the trap. aoqi@0: bool found_trap = false; aoqi@0: for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { aoqi@0: Node* obs = not_ctl->fast_out(j); aoqi@0: if (obs->in(0) == not_ctl && obs->is_Call() && aoqi@0: (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { aoqi@0: found_trap = true; break; aoqi@0: } aoqi@0: } aoqi@0: if (found_trap) { aoqi@0: ctl = iff->in(0); // This test feeds a harmless uncommon trap. aoqi@0: continue; aoqi@0: } aoqi@0: } aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: // If we get this far, we have an allocation which immediately aoqi@0: // precedes the arraycopy, and we can take over zeroing the new object. aoqi@0: // The arraycopy will finish the initialization, and provide aoqi@0: // a new control state to which we will anchor the destination pointer. aoqi@0: aoqi@0: return alloc; aoqi@0: } aoqi@0: aoqi@0: // Helper for initialization of arrays, creating a ClearArray. aoqi@0: // It writes zero bits in [start..end), within the body of an array object. aoqi@0: // The memory effects are all chained onto the 'adr_type' alias category. aoqi@0: // aoqi@0: // Since the object is otherwise uninitialized, we are free aoqi@0: // to put a little "slop" around the edges of the cleared area, aoqi@0: // as long as it does not go back into the array's header, aoqi@0: // or beyond the array end within the heap. aoqi@0: // aoqi@0: // The lower edge can be rounded down to the nearest jint and the aoqi@0: // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. aoqi@0: // aoqi@0: // Arguments: aoqi@0: // adr_type memory slice where writes are generated aoqi@0: // dest oop of the destination array aoqi@0: // basic_elem_type element type of the destination aoqi@0: // slice_idx array index of first element to store aoqi@0: // slice_len number of elements to store (or NULL) aoqi@0: // dest_size total size in bytes of the array object aoqi@0: // aoqi@0: // Exactly one of slice_len or dest_size must be non-NULL. aoqi@0: // If dest_size is non-NULL, zeroing extends to the end of the object. aoqi@0: // If slice_len is non-NULL, the slice_idx value must be a constant. aoqi@0: void aoqi@0: LibraryCallKit::generate_clear_array(const TypePtr* adr_type, aoqi@0: Node* dest, aoqi@0: BasicType basic_elem_type, aoqi@0: Node* slice_idx, aoqi@0: Node* slice_len, aoqi@0: Node* dest_size) { aoqi@0: // one or the other but not both of slice_len and dest_size: aoqi@0: assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); aoqi@0: if (slice_len == NULL) slice_len = top(); aoqi@0: if (dest_size == NULL) dest_size = top(); aoqi@0: aoqi@0: // operate on this memory slice: aoqi@0: Node* mem = memory(adr_type); // memory slice to operate on aoqi@0: aoqi@0: // scaling and rounding of indexes: aoqi@0: int scale = exact_log2(type2aelembytes(basic_elem_type)); aoqi@0: int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); aoqi@0: int clear_low = (-1 << scale) & (BytesPerInt - 1); aoqi@0: int bump_bit = (-1 << scale) & BytesPerInt; aoqi@0: aoqi@0: // determine constant starts and ends aoqi@0: const intptr_t BIG_NEG = -128; aoqi@0: assert(BIG_NEG + 2*abase < 0, "neg enough"); aoqi@0: intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); aoqi@0: intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); aoqi@0: if (slice_len_con == 0) { aoqi@0: return; // nothing to do here aoqi@0: } aoqi@0: intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; aoqi@0: intptr_t end_con = find_intptr_t_con(dest_size, -1); aoqi@0: if (slice_idx_con >= 0 && slice_len_con >= 0) { aoqi@0: assert(end_con < 0, "not two cons"); aoqi@0: end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), aoqi@0: BytesPerLong); aoqi@0: } aoqi@0: aoqi@0: if (start_con >= 0 && end_con >= 0) { aoqi@0: // Constant start and end. Simple. aoqi@0: mem = ClearArrayNode::clear_memory(control(), mem, dest, aoqi@0: start_con, end_con, &_gvn); aoqi@0: } else if (start_con >= 0 && dest_size != top()) { aoqi@0: // Constant start, pre-rounded end after the tail of the array. aoqi@0: Node* end = dest_size; aoqi@0: mem = ClearArrayNode::clear_memory(control(), mem, dest, aoqi@0: start_con, end, &_gvn); aoqi@0: } else if (start_con >= 0 && slice_len != top()) { aoqi@0: // Constant start, non-constant end. End needs rounding up. aoqi@0: // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) aoqi@0: intptr_t end_base = abase + (slice_idx_con << scale); aoqi@0: int end_round = (-1 << scale) & (BytesPerLong - 1); aoqi@0: Node* end = ConvI2X(slice_len); aoqi@0: if (scale != 0) aoqi@0: end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) )); aoqi@0: end_base += end_round; aoqi@0: end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base))); aoqi@0: end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round))); aoqi@0: mem = ClearArrayNode::clear_memory(control(), mem, dest, aoqi@0: start_con, end, &_gvn); aoqi@0: } else if (start_con < 0 && dest_size != top()) { aoqi@0: // Non-constant start, pre-rounded end after the tail of the array. aoqi@0: // This is almost certainly a "round-to-end" operation. aoqi@0: Node* start = slice_idx; aoqi@0: start = ConvI2X(start); aoqi@0: if (scale != 0) aoqi@0: start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) )); aoqi@0: start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase))); aoqi@0: if ((bump_bit | clear_low) != 0) { aoqi@0: int to_clear = (bump_bit | clear_low); aoqi@0: // Align up mod 8, then store a jint zero unconditionally aoqi@0: // just before the mod-8 boundary. aoqi@0: if (((abase + bump_bit) & ~to_clear) - bump_bit aoqi@0: < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { aoqi@0: bump_bit = 0; aoqi@0: assert((abase & to_clear) == 0, "array base must be long-aligned"); aoqi@0: } else { aoqi@0: // Bump 'start' up to (or past) the next jint boundary: aoqi@0: start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit))); aoqi@0: assert((abase & clear_low) == 0, "array base must be int-aligned"); aoqi@0: } aoqi@0: // Round bumped 'start' down to jlong boundary in body of array. aoqi@0: start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear))); aoqi@0: if (bump_bit != 0) { aoqi@0: // Store a zero to the immediately preceding jint: aoqi@0: Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit))); aoqi@0: Node* p1 = basic_plus_adr(dest, x1); aoqi@0: mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered); aoqi@0: mem = _gvn.transform(mem); aoqi@0: } aoqi@0: } aoqi@0: Node* end = dest_size; // pre-rounded aoqi@0: mem = ClearArrayNode::clear_memory(control(), mem, dest, aoqi@0: start, end, &_gvn); aoqi@0: } else { aoqi@0: // Non-constant start, unrounded non-constant end. aoqi@0: // (Nobody zeroes a random midsection of an array using this routine.) aoqi@0: ShouldNotReachHere(); // fix caller aoqi@0: } aoqi@0: aoqi@0: // Done. aoqi@0: set_memory(mem, adr_type); aoqi@0: } aoqi@0: aoqi@0: aoqi@0: bool aoqi@0: LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, aoqi@0: BasicType basic_elem_type, aoqi@0: AllocateNode* alloc, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* dest_size, bool dest_uninitialized) { aoqi@0: // See if there is an advantage from block transfer. aoqi@0: int scale = exact_log2(type2aelembytes(basic_elem_type)); aoqi@0: if (scale >= LogBytesPerLong) aoqi@0: return false; // it is already a block transfer aoqi@0: aoqi@0: // Look at the alignment of the starting offsets. aoqi@0: int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); aoqi@0: aoqi@0: intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); aoqi@0: intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); aoqi@0: if (src_off_con < 0 || dest_off_con < 0) aoqi@0: // At present, we can only understand constants. aoqi@0: return false; aoqi@0: aoqi@0: intptr_t src_off = abase + (src_off_con << scale); aoqi@0: intptr_t dest_off = abase + (dest_off_con << scale); aoqi@0: aoqi@0: if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { aoqi@0: // Non-aligned; too bad. aoqi@0: // One more chance: Pick off an initial 32-bit word. aoqi@0: // This is a common case, since abase can be odd mod 8. aoqi@0: if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && aoqi@0: ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { aoqi@0: Node* sptr = basic_plus_adr(src, src_off); aoqi@0: Node* dptr = basic_plus_adr(dest, dest_off); aoqi@0: Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); aoqi@0: store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered); aoqi@0: src_off += BytesPerInt; aoqi@0: dest_off += BytesPerInt; aoqi@0: } else { aoqi@0: return false; aoqi@0: } aoqi@0: } aoqi@0: assert(src_off % BytesPerLong == 0, ""); aoqi@0: assert(dest_off % BytesPerLong == 0, ""); aoqi@0: aoqi@0: // Do this copy by giant steps. aoqi@0: Node* sptr = basic_plus_adr(src, src_off); aoqi@0: Node* dptr = basic_plus_adr(dest, dest_off); aoqi@0: Node* countx = dest_size; aoqi@0: countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off))); aoqi@0: countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong))); aoqi@0: aoqi@0: bool disjoint_bases = true; // since alloc != NULL aoqi@0: generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, aoqi@0: sptr, NULL, dptr, NULL, countx, dest_uninitialized); aoqi@0: aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: aoqi@0: // Helper function; generates code for the slow case. aoqi@0: // We make a call to a runtime method which emulates the native method, aoqi@0: // but without the native wrapper overhead. aoqi@0: void aoqi@0: LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* copy_length, bool dest_uninitialized) { aoqi@0: assert(!dest_uninitialized, "Invariant"); aoqi@0: Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, aoqi@0: OptoRuntime::slow_arraycopy_Type(), aoqi@0: OptoRuntime::slow_arraycopy_Java(), aoqi@0: "slow_arraycopy", adr_type, aoqi@0: src, src_offset, dest, dest_offset, aoqi@0: copy_length); aoqi@0: aoqi@0: // Handle exceptions thrown by this fellow: aoqi@0: make_slow_call_ex(call, env()->Throwable_klass(), false); aoqi@0: } aoqi@0: aoqi@0: // Helper function; generates code for cases requiring runtime checks. aoqi@0: Node* aoqi@0: LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, aoqi@0: Node* dest_elem_klass, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* copy_length, bool dest_uninitialized) { aoqi@0: if (stopped()) return NULL; aoqi@0: aoqi@0: address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); aoqi@0: if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: // Pick out the parameters required to perform a store-check aoqi@0: // for the target array. This is an optimistic check. It will aoqi@0: // look in each non-null element's class, at the desired klass's aoqi@0: // super_check_offset, for the desired klass. aoqi@0: int sco_offset = in_bytes(Klass::super_check_offset_offset()); aoqi@0: Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); aoqi@0: Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered); aoqi@0: Node* check_offset = ConvI2X(_gvn.transform(n3)); aoqi@0: Node* check_value = dest_elem_klass; aoqi@0: aoqi@0: Node* src_start = array_element_address(src, src_offset, T_OBJECT); aoqi@0: Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); aoqi@0: aoqi@0: // (We know the arrays are never conjoint, because their types differ.) aoqi@0: Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, aoqi@0: OptoRuntime::checkcast_arraycopy_Type(), aoqi@0: copyfunc_addr, "checkcast_arraycopy", adr_type, aoqi@0: // five arguments, of which two are aoqi@0: // intptr_t (jlong in LP64) aoqi@0: src_start, dest_start, aoqi@0: copy_length XTOP, aoqi@0: check_offset XTOP, aoqi@0: check_value); aoqi@0: aoqi@0: return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); aoqi@0: } aoqi@0: aoqi@0: aoqi@0: // Helper function; generates code for cases requiring runtime checks. aoqi@0: Node* aoqi@0: LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* copy_length, bool dest_uninitialized) { aoqi@0: assert(!dest_uninitialized, "Invariant"); aoqi@0: if (stopped()) return NULL; aoqi@0: address copyfunc_addr = StubRoutines::generic_arraycopy(); aoqi@0: if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. aoqi@0: return NULL; aoqi@0: } aoqi@0: aoqi@0: Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, aoqi@0: OptoRuntime::generic_arraycopy_Type(), aoqi@0: copyfunc_addr, "generic_arraycopy", adr_type, aoqi@0: src, src_offset, dest, dest_offset, copy_length); aoqi@0: aoqi@0: return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); aoqi@0: } aoqi@0: aoqi@0: // Helper function; generates the fast out-of-line call to an arraycopy stub. aoqi@0: void aoqi@0: LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, aoqi@0: BasicType basic_elem_type, aoqi@0: bool disjoint_bases, aoqi@0: Node* src, Node* src_offset, aoqi@0: Node* dest, Node* dest_offset, aoqi@0: Node* copy_length, bool dest_uninitialized) { aoqi@0: if (stopped()) return; // nothing to do aoqi@0: aoqi@0: Node* src_start = src; aoqi@0: Node* dest_start = dest; aoqi@0: if (src_offset != NULL || dest_offset != NULL) { aoqi@0: assert(src_offset != NULL && dest_offset != NULL, ""); aoqi@0: src_start = array_element_address(src, src_offset, basic_elem_type); aoqi@0: dest_start = array_element_address(dest, dest_offset, basic_elem_type); aoqi@0: } aoqi@0: aoqi@0: // Figure out which arraycopy runtime method to call. aoqi@0: const char* copyfunc_name = "arraycopy"; aoqi@0: address copyfunc_addr = aoqi@0: basictype2arraycopy(basic_elem_type, src_offset, dest_offset, aoqi@0: disjoint_bases, copyfunc_name, dest_uninitialized); aoqi@0: aoqi@0: // Call it. Note that the count_ix value is not scaled to a byte-size. aoqi@0: make_runtime_call(RC_LEAF|RC_NO_FP, aoqi@0: OptoRuntime::fast_arraycopy_Type(), aoqi@0: copyfunc_addr, copyfunc_name, adr_type, aoqi@0: src_start, dest_start, copy_length XTOP); aoqi@0: } aoqi@0: aoqi@0: //-------------inline_encodeISOArray----------------------------------- aoqi@0: // encode char[] to byte[] in ISO_8859_1 aoqi@0: bool LibraryCallKit::inline_encodeISOArray() { aoqi@0: assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); aoqi@0: // no receiver since it is static method aoqi@0: Node *src = argument(0); aoqi@0: Node *src_offset = argument(1); aoqi@0: Node *dst = argument(2); aoqi@0: Node *dst_offset = argument(3); aoqi@0: Node *length = argument(4); aoqi@0: aoqi@0: const Type* src_type = src->Value(&_gvn); aoqi@0: const Type* dst_type = dst->Value(&_gvn); aoqi@0: const TypeAryPtr* top_src = src_type->isa_aryptr(); aoqi@0: const TypeAryPtr* top_dest = dst_type->isa_aryptr(); aoqi@0: if (top_src == NULL || top_src->klass() == NULL || aoqi@0: top_dest == NULL || top_dest->klass() == NULL) { aoqi@0: // failed array check aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: // Figure out the size and type of the elements we will be copying. aoqi@0: BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); aoqi@0: BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); aoqi@0: if (src_elem != T_CHAR || dst_elem != T_BYTE) { aoqi@0: return false; aoqi@0: } aoqi@0: Node* src_start = array_element_address(src, src_offset, src_elem); aoqi@0: Node* dst_start = array_element_address(dst, dst_offset, dst_elem); aoqi@0: // 'src_start' points to src array + scaled offset aoqi@0: // 'dst_start' points to dst array + scaled offset aoqi@0: aoqi@0: const TypeAryPtr* mtype = TypeAryPtr::BYTES; aoqi@0: Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); aoqi@0: enc = _gvn.transform(enc); aoqi@0: Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc)); aoqi@0: set_memory(res_mem, mtype); aoqi@0: set_result(enc); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: /** aoqi@0: * Calculate CRC32 for byte. aoqi@0: * int java.util.zip.CRC32.update(int crc, int b) aoqi@0: */ aoqi@0: bool LibraryCallKit::inline_updateCRC32() { aoqi@0: assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); aoqi@0: assert(callee()->signature()->size() == 2, "update has 2 parameters"); aoqi@0: // no receiver since it is static method aoqi@0: Node* crc = argument(0); // type: int aoqi@0: Node* b = argument(1); // type: int aoqi@0: aoqi@0: /* aoqi@0: * int c = ~ crc; aoqi@0: * b = timesXtoThe32[(b ^ c) & 0xFF]; aoqi@0: * b = b ^ (c >>> 8); aoqi@0: * crc = ~b; aoqi@0: */ aoqi@0: aoqi@0: Node* M1 = intcon(-1); aoqi@0: crc = _gvn.transform(new (C) XorINode(crc, M1)); aoqi@0: Node* result = _gvn.transform(new (C) XorINode(crc, b)); aoqi@0: result = _gvn.transform(new (C) AndINode(result, intcon(0xFF))); aoqi@0: aoqi@0: Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); aoqi@0: Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2))); aoqi@0: Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); aoqi@0: result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); aoqi@0: aoqi@0: crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8))); aoqi@0: result = _gvn.transform(new (C) XorINode(crc, result)); aoqi@0: result = _gvn.transform(new (C) XorINode(result, M1)); aoqi@0: set_result(result); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: /** aoqi@0: * Calculate CRC32 for byte[] array. aoqi@0: * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) aoqi@0: */ aoqi@0: bool LibraryCallKit::inline_updateBytesCRC32() { aoqi@0: assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); aoqi@0: assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); aoqi@0: // no receiver since it is static method aoqi@0: Node* crc = argument(0); // type: int aoqi@0: Node* src = argument(1); // type: oop aoqi@0: Node* offset = argument(2); // type: int aoqi@0: Node* length = argument(3); // type: int aoqi@0: aoqi@0: const Type* src_type = src->Value(&_gvn); aoqi@0: const TypeAryPtr* top_src = src_type->isa_aryptr(); aoqi@0: if (top_src == NULL || top_src->klass() == NULL) { aoqi@0: // failed array check aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: // Figure out the size and type of the elements we will be copying. aoqi@0: BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); aoqi@0: if (src_elem != T_BYTE) { aoqi@0: return false; aoqi@0: } aoqi@0: aoqi@0: // 'src_start' points to src array + scaled offset aoqi@0: Node* src_start = array_element_address(src, offset, src_elem); aoqi@0: aoqi@0: // We assume that range check is done by caller. aoqi@0: // TODO: generate range check (offset+length < src.length) in debug VM. aoqi@0: aoqi@0: // Call the stub. aoqi@0: address stubAddr = StubRoutines::updateBytesCRC32(); aoqi@0: const char *stubName = "updateBytesCRC32"; aoqi@0: aoqi@0: Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), aoqi@0: stubAddr, stubName, TypePtr::BOTTOM, aoqi@0: crc, src_start, length); aoqi@0: Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); aoqi@0: set_result(result); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: /** aoqi@0: * Calculate CRC32 for ByteBuffer. aoqi@0: * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) aoqi@0: */ aoqi@0: bool LibraryCallKit::inline_updateByteBufferCRC32() { aoqi@0: assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); aoqi@0: assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); aoqi@0: // no receiver since it is static method aoqi@0: Node* crc = argument(0); // type: int aoqi@0: Node* src = argument(1); // type: long aoqi@0: Node* offset = argument(3); // type: int aoqi@0: Node* length = argument(4); // type: int aoqi@0: aoqi@0: src = ConvL2X(src); // adjust Java long to machine word aoqi@0: Node* base = _gvn.transform(new (C) CastX2PNode(src)); aoqi@0: offset = ConvI2X(offset); aoqi@0: aoqi@0: // 'src_start' points to src array + scaled offset aoqi@0: Node* src_start = basic_plus_adr(top(), base, offset); aoqi@0: aoqi@0: // Call the stub. aoqi@0: address stubAddr = StubRoutines::updateBytesCRC32(); aoqi@0: const char *stubName = "updateBytesCRC32"; aoqi@0: aoqi@0: Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), aoqi@0: stubAddr, stubName, TypePtr::BOTTOM, aoqi@0: crc, src_start, length); aoqi@0: Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); aoqi@0: set_result(result); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //----------------------------inline_reference_get---------------------------- aoqi@0: // public T java.lang.ref.Reference.get(); aoqi@0: bool LibraryCallKit::inline_reference_get() { aoqi@0: const int referent_offset = java_lang_ref_Reference::referent_offset; aoqi@0: guarantee(referent_offset > 0, "should have already been set"); aoqi@0: aoqi@0: // Get the argument: aoqi@0: Node* reference_obj = null_check_receiver(); aoqi@0: if (stopped()) return true; aoqi@0: aoqi@0: Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); aoqi@0: aoqi@0: ciInstanceKlass* klass = env()->Object_klass(); aoqi@0: const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); aoqi@0: aoqi@0: Node* no_ctrl = NULL; aoqi@0: Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); aoqi@0: aoqi@0: // Use the pre-barrier to record the value in the referent field aoqi@0: pre_barrier(false /* do_load */, aoqi@0: control(), aoqi@0: NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, aoqi@0: result /* pre_val */, aoqi@0: T_OBJECT); aoqi@0: aoqi@0: // Add memory barrier to prevent commoning reads from this field aoqi@0: // across safepoint since GC can change its value. aoqi@0: insert_mem_bar(Op_MemBarCPUOrder); aoqi@0: aoqi@0: set_result(result); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: aoqi@0: Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, aoqi@0: bool is_exact=true, bool is_static=false) { aoqi@0: aoqi@0: const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); aoqi@0: assert(tinst != NULL, "obj is null"); aoqi@0: assert(tinst->klass()->is_loaded(), "obj is not loaded"); aoqi@0: assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); aoqi@0: aoqi@0: ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), aoqi@0: ciSymbol::make(fieldTypeString), aoqi@0: is_static); aoqi@0: if (field == NULL) return (Node *) NULL; aoqi@0: assert (field != NULL, "undefined field"); aoqi@0: aoqi@0: // Next code copied from Parse::do_get_xxx(): aoqi@0: aoqi@0: // Compute address and memory type. aoqi@0: int offset = field->offset_in_bytes(); aoqi@0: bool is_vol = field->is_volatile(); aoqi@0: ciType* field_klass = field->type(); aoqi@0: assert(field_klass->is_loaded(), "should be loaded"); aoqi@0: const TypePtr* adr_type = C->alias_type(field)->adr_type(); aoqi@0: Node *adr = basic_plus_adr(fromObj, fromObj, offset); aoqi@0: BasicType bt = field->layout_type(); aoqi@0: aoqi@0: // Build the resultant type of the load aoqi@0: const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass()); aoqi@0: aoqi@0: // Build the load. aoqi@0: Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol); aoqi@0: return loadedField; aoqi@0: } aoqi@0: aoqi@0: aoqi@0: //------------------------------inline_aescrypt_Block----------------------- aoqi@0: bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { aoqi@0: address stubAddr; aoqi@0: const char *stubName; aoqi@0: assert(UseAES, "need AES instruction support"); aoqi@0: aoqi@0: switch(id) { aoqi@0: case vmIntrinsics::_aescrypt_encryptBlock: aoqi@0: stubAddr = StubRoutines::aescrypt_encryptBlock(); aoqi@0: stubName = "aescrypt_encryptBlock"; aoqi@0: break; aoqi@0: case vmIntrinsics::_aescrypt_decryptBlock: aoqi@0: stubAddr = StubRoutines::aescrypt_decryptBlock(); aoqi@0: stubName = "aescrypt_decryptBlock"; aoqi@0: break; aoqi@0: } aoqi@0: if (stubAddr == NULL) return false; aoqi@0: aoqi@0: Node* aescrypt_object = argument(0); aoqi@0: Node* src = argument(1); aoqi@0: Node* src_offset = argument(2); aoqi@0: Node* dest = argument(3); aoqi@0: Node* dest_offset = argument(4); aoqi@0: aoqi@0: // (1) src and dest are arrays. aoqi@0: const Type* src_type = src->Value(&_gvn); aoqi@0: const Type* dest_type = dest->Value(&_gvn); aoqi@0: const TypeAryPtr* top_src = src_type->isa_aryptr(); aoqi@0: const TypeAryPtr* top_dest = dest_type->isa_aryptr(); aoqi@0: assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); aoqi@0: aoqi@0: // for the quick and dirty code we will skip all the checks. aoqi@0: // we are just trying to get the call to be generated. aoqi@0: Node* src_start = src; aoqi@0: Node* dest_start = dest; aoqi@0: if (src_offset != NULL || dest_offset != NULL) { aoqi@0: assert(src_offset != NULL && dest_offset != NULL, ""); aoqi@0: src_start = array_element_address(src, src_offset, T_BYTE); aoqi@0: dest_start = array_element_address(dest, dest_offset, T_BYTE); aoqi@0: } aoqi@0: aoqi@0: // now need to get the start of its expanded key array aoqi@0: // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java aoqi@0: Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); aoqi@0: if (k_start == NULL) return false; aoqi@0: aoqi@0: if (Matcher::pass_original_key_for_aes()) { aoqi@0: // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to aoqi@0: // compatibility issues between Java key expansion and SPARC crypto instructions aoqi@0: Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); aoqi@0: if (original_k_start == NULL) return false; aoqi@0: aoqi@0: // Call the stub. aoqi@0: make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), aoqi@0: stubAddr, stubName, TypePtr::BOTTOM, aoqi@0: src_start, dest_start, k_start, original_k_start); aoqi@0: } else { aoqi@0: // Call the stub. aoqi@0: make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), aoqi@0: stubAddr, stubName, TypePtr::BOTTOM, aoqi@0: src_start, dest_start, k_start); aoqi@0: } aoqi@0: aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- aoqi@0: bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { aoqi@0: address stubAddr; aoqi@0: const char *stubName; aoqi@0: aoqi@0: assert(UseAES, "need AES instruction support"); aoqi@0: aoqi@0: switch(id) { aoqi@0: case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: aoqi@0: stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); aoqi@0: stubName = "cipherBlockChaining_encryptAESCrypt"; aoqi@0: break; aoqi@0: case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: aoqi@0: stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); aoqi@0: stubName = "cipherBlockChaining_decryptAESCrypt"; aoqi@0: break; aoqi@0: } aoqi@0: if (stubAddr == NULL) return false; aoqi@0: aoqi@0: Node* cipherBlockChaining_object = argument(0); aoqi@0: Node* src = argument(1); aoqi@0: Node* src_offset = argument(2); aoqi@0: Node* len = argument(3); aoqi@0: Node* dest = argument(4); aoqi@0: Node* dest_offset = argument(5); aoqi@0: aoqi@0: // (1) src and dest are arrays. aoqi@0: const Type* src_type = src->Value(&_gvn); aoqi@0: const Type* dest_type = dest->Value(&_gvn); aoqi@0: const TypeAryPtr* top_src = src_type->isa_aryptr(); aoqi@0: const TypeAryPtr* top_dest = dest_type->isa_aryptr(); aoqi@0: assert (top_src != NULL && top_src->klass() != NULL aoqi@0: && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); aoqi@0: aoqi@0: // checks are the responsibility of the caller aoqi@0: Node* src_start = src; aoqi@0: Node* dest_start = dest; aoqi@0: if (src_offset != NULL || dest_offset != NULL) { aoqi@0: assert(src_offset != NULL && dest_offset != NULL, ""); aoqi@0: src_start = array_element_address(src, src_offset, T_BYTE); aoqi@0: dest_start = array_element_address(dest, dest_offset, T_BYTE); aoqi@0: } aoqi@0: aoqi@0: // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object aoqi@0: // (because of the predicated logic executed earlier). aoqi@0: // so we cast it here safely. aoqi@0: // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java aoqi@0: aoqi@0: Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); aoqi@0: if (embeddedCipherObj == NULL) return false; aoqi@0: aoqi@0: // cast it to what we know it will be at runtime aoqi@0: const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); aoqi@0: assert(tinst != NULL, "CBC obj is null"); aoqi@0: assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); aoqi@0: ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); aoqi@0: if (!klass_AESCrypt->is_loaded()) return false; aoqi@0: aoqi@0: ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); aoqi@0: const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); aoqi@0: const TypeOopPtr* xtype = aklass->as_instance_type(); aoqi@0: Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype); aoqi@0: aescrypt_object = _gvn.transform(aescrypt_object); aoqi@0: aoqi@0: // we need to get the start of the aescrypt_object's expanded key array aoqi@0: Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); aoqi@0: if (k_start == NULL) return false; aoqi@0: aoqi@0: // similarly, get the start address of the r vector aoqi@0: Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); aoqi@0: if (objRvec == NULL) return false; aoqi@0: Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); aoqi@0: aoqi@0: Node* cbcCrypt; aoqi@0: if (Matcher::pass_original_key_for_aes()) { aoqi@0: // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to aoqi@0: // compatibility issues between Java key expansion and SPARC crypto instructions aoqi@0: Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); aoqi@0: if (original_k_start == NULL) return false; aoqi@0: aoqi@0: // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start aoqi@0: cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, aoqi@0: OptoRuntime::cipherBlockChaining_aescrypt_Type(), aoqi@0: stubAddr, stubName, TypePtr::BOTTOM, aoqi@0: src_start, dest_start, k_start, r_start, len, original_k_start); aoqi@0: } else { aoqi@0: // Call the stub, passing src_start, dest_start, k_start, r_start and src_len aoqi@0: cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, aoqi@0: OptoRuntime::cipherBlockChaining_aescrypt_Type(), aoqi@0: stubAddr, stubName, TypePtr::BOTTOM, aoqi@0: src_start, dest_start, k_start, r_start, len); aoqi@0: } aoqi@0: aoqi@0: // return cipher length (int) aoqi@0: Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms)); aoqi@0: set_result(retvalue); aoqi@0: return true; aoqi@0: } aoqi@0: aoqi@0: //------------------------------get_key_start_from_aescrypt_object----------------------- aoqi@0: Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { aoqi@0: Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); aoqi@0: assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); aoqi@0: if (objAESCryptKey == NULL) return (Node *) NULL; aoqi@0: aoqi@0: // now have the array, need to get the start address of the K array aoqi@0: Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); aoqi@0: return k_start; aoqi@0: } aoqi@0: aoqi@0: //------------------------------get_original_key_start_from_aescrypt_object----------------------- aoqi@0: Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { aoqi@0: Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); aoqi@0: assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); aoqi@0: if (objAESCryptKey == NULL) return (Node *) NULL; aoqi@0: aoqi@0: // now have the array, need to get the start address of the lastKey array aoqi@0: Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); aoqi@0: return original_k_start; aoqi@0: } aoqi@0: aoqi@0: //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- aoqi@0: // Return node representing slow path of predicate check. aoqi@0: // the pseudo code we want to emulate with this predicate is: aoqi@0: // for encryption: aoqi@0: // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath aoqi@0: // for decryption: aoqi@0: // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath aoqi@0: // note cipher==plain is more conservative than the original java code but that's OK aoqi@0: // aoqi@0: Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { aoqi@0: // First, check receiver for NULL since it is virtual method. aoqi@0: Node* objCBC = argument(0); aoqi@0: objCBC = null_check(objCBC); aoqi@0: aoqi@0: if (stopped()) return NULL; // Always NULL aoqi@0: aoqi@0: // Load embeddedCipher field of CipherBlockChaining object. aoqi@0: Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); aoqi@0: aoqi@0: // get AESCrypt klass for instanceOf check aoqi@0: // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point aoqi@0: // will have same classloader as CipherBlockChaining object aoqi@0: const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); aoqi@0: assert(tinst != NULL, "CBCobj is null"); aoqi@0: assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); aoqi@0: aoqi@0: // we want to do an instanceof comparison against the AESCrypt class aoqi@0: ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); aoqi@0: if (!klass_AESCrypt->is_loaded()) { aoqi@0: // if AESCrypt is not even loaded, we never take the intrinsic fast path aoqi@0: Node* ctrl = control(); aoqi@0: set_control(top()); // no regular fast path aoqi@0: return ctrl; aoqi@0: } aoqi@0: ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); aoqi@0: aoqi@0: Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); aoqi@0: Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1))); aoqi@0: Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne)); aoqi@0: aoqi@0: Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); aoqi@0: aoqi@0: // for encryption, we are done aoqi@0: if (!decrypting) aoqi@0: return instof_false; // even if it is NULL aoqi@0: aoqi@0: // for decryption, we need to add a further check to avoid aoqi@0: // taking the intrinsic path when cipher and plain are the same aoqi@0: // see the original java code for why. aoqi@0: RegionNode* region = new(C) RegionNode(3); aoqi@0: region->init_req(1, instof_false); aoqi@0: Node* src = argument(1); aoqi@0: Node* dest = argument(4); aoqi@0: Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest)); aoqi@0: Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq)); aoqi@0: Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); aoqi@0: region->init_req(2, src_dest_conjoint); aoqi@0: aoqi@0: record_for_igvn(region); aoqi@0: return _gvn.transform(region); aoqi@0: }