duke@435: /* xdono@631: * Copyright 1999-2008 Sun Microsystems, Inc. All Rights Reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * duke@435: * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, duke@435: * CA 95054 USA or visit www.sun.com if you need additional information or duke@435: * have any questions. duke@435: * duke@435: */ duke@435: duke@435: #include "incls/_precompiled.incl" duke@435: #include "incls/_library_call.cpp.incl" duke@435: duke@435: class LibraryIntrinsic : public InlineCallGenerator { duke@435: // Extend the set of intrinsics known to the runtime: duke@435: public: duke@435: private: duke@435: bool _is_virtual; duke@435: vmIntrinsics::ID _intrinsic_id; duke@435: duke@435: public: duke@435: LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id) duke@435: : InlineCallGenerator(m), duke@435: _is_virtual(is_virtual), duke@435: _intrinsic_id(id) duke@435: { duke@435: } duke@435: virtual bool is_intrinsic() const { return true; } duke@435: virtual bool is_virtual() const { return _is_virtual; } duke@435: virtual JVMState* generate(JVMState* jvms); duke@435: vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } duke@435: }; duke@435: duke@435: duke@435: // Local helper class for LibraryIntrinsic: duke@435: class LibraryCallKit : public GraphKit { duke@435: private: duke@435: LibraryIntrinsic* _intrinsic; // the library intrinsic being called duke@435: duke@435: public: duke@435: LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic) duke@435: : GraphKit(caller), duke@435: _intrinsic(intrinsic) duke@435: { duke@435: } duke@435: duke@435: ciMethod* caller() const { return jvms()->method(); } duke@435: int bci() const { return jvms()->bci(); } duke@435: LibraryIntrinsic* intrinsic() const { return _intrinsic; } duke@435: vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } duke@435: ciMethod* callee() const { return _intrinsic->method(); } duke@435: ciSignature* signature() const { return callee()->signature(); } duke@435: int arg_size() const { return callee()->arg_size(); } duke@435: duke@435: bool try_to_inline(); duke@435: duke@435: // Helper functions to inline natives duke@435: void push_result(RegionNode* region, PhiNode* value); duke@435: Node* generate_guard(Node* test, RegionNode* region, float true_prob); duke@435: Node* generate_slow_guard(Node* test, RegionNode* region); duke@435: Node* generate_fair_guard(Node* test, RegionNode* region); duke@435: Node* generate_negative_guard(Node* index, RegionNode* region, duke@435: // resulting CastII of index: duke@435: Node* *pos_index = NULL); duke@435: Node* generate_nonpositive_guard(Node* index, bool never_negative, duke@435: // resulting CastII of index: duke@435: Node* *pos_index = NULL); duke@435: Node* generate_limit_guard(Node* offset, Node* subseq_length, duke@435: Node* array_length, duke@435: RegionNode* region); duke@435: Node* generate_current_thread(Node* &tls_output); duke@435: address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, duke@435: bool disjoint_bases, const char* &name); duke@435: Node* load_mirror_from_klass(Node* klass); duke@435: Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, duke@435: int nargs, duke@435: RegionNode* region, int null_path, duke@435: int offset); duke@435: Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs, duke@435: RegionNode* region, int null_path) { duke@435: int offset = java_lang_Class::klass_offset_in_bytes(); duke@435: return load_klass_from_mirror_common(mirror, never_see_null, nargs, duke@435: region, null_path, duke@435: offset); duke@435: } duke@435: Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, duke@435: int nargs, duke@435: RegionNode* region, int null_path) { duke@435: int offset = java_lang_Class::array_klass_offset_in_bytes(); duke@435: return load_klass_from_mirror_common(mirror, never_see_null, nargs, duke@435: region, null_path, duke@435: offset); duke@435: } duke@435: Node* generate_access_flags_guard(Node* kls, duke@435: int modifier_mask, int modifier_bits, duke@435: RegionNode* region); duke@435: Node* generate_interface_guard(Node* kls, RegionNode* region); duke@435: Node* generate_array_guard(Node* kls, RegionNode* region) { duke@435: return generate_array_guard_common(kls, region, false, false); duke@435: } duke@435: Node* generate_non_array_guard(Node* kls, RegionNode* region) { duke@435: return generate_array_guard_common(kls, region, false, true); duke@435: } duke@435: Node* generate_objArray_guard(Node* kls, RegionNode* region) { duke@435: return generate_array_guard_common(kls, region, true, false); duke@435: } duke@435: Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { duke@435: return generate_array_guard_common(kls, region, true, true); duke@435: } duke@435: Node* generate_array_guard_common(Node* kls, RegionNode* region, duke@435: bool obj_array, bool not_array); duke@435: Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); duke@435: CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, duke@435: bool is_virtual = false, bool is_static = false); duke@435: CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { duke@435: return generate_method_call(method_id, false, true); duke@435: } duke@435: CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { duke@435: return generate_method_call(method_id, true, false); duke@435: } duke@435: duke@435: bool inline_string_compareTo(); duke@435: bool inline_string_indexOf(); duke@435: Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); duke@435: Node* pop_math_arg(); duke@435: bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); duke@435: bool inline_math_native(vmIntrinsics::ID id); duke@435: bool inline_trig(vmIntrinsics::ID id); duke@435: bool inline_trans(vmIntrinsics::ID id); duke@435: bool inline_abs(vmIntrinsics::ID id); duke@435: bool inline_sqrt(vmIntrinsics::ID id); duke@435: bool inline_pow(vmIntrinsics::ID id); duke@435: bool inline_exp(vmIntrinsics::ID id); duke@435: bool inline_min_max(vmIntrinsics::ID id); duke@435: Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); duke@435: // This returns Type::AnyPtr, RawPtr, or OopPtr. duke@435: int classify_unsafe_addr(Node* &base, Node* &offset); duke@435: Node* make_unsafe_address(Node* base, Node* offset); duke@435: bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); duke@435: bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); duke@435: bool inline_unsafe_allocate(); duke@435: bool inline_unsafe_copyMemory(); duke@435: bool inline_native_currentThread(); duke@435: bool inline_native_time_funcs(bool isNano); duke@435: bool inline_native_isInterrupted(); duke@435: bool inline_native_Class_query(vmIntrinsics::ID id); duke@435: bool inline_native_subtype_check(); duke@435: duke@435: bool inline_native_newArray(); duke@435: bool inline_native_getLength(); duke@435: bool inline_array_copyOf(bool is_copyOfRange); rasbold@604: bool inline_array_equals(); duke@435: bool inline_native_clone(bool is_virtual); duke@435: bool inline_native_Reflection_getCallerClass(); duke@435: bool inline_native_AtomicLong_get(); duke@435: bool inline_native_AtomicLong_attemptUpdate(); duke@435: bool is_method_invoke_or_aux_frame(JVMState* jvms); duke@435: // Helper function for inlining native object hash method duke@435: bool inline_native_hashcode(bool is_virtual, bool is_static); duke@435: bool inline_native_getClass(); duke@435: duke@435: // Helper functions for inlining arraycopy duke@435: bool inline_arraycopy(); duke@435: void generate_arraycopy(const TypePtr* adr_type, duke@435: BasicType basic_elem_type, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* copy_length, duke@435: int nargs, // arguments on stack for debug info duke@435: bool disjoint_bases = false, duke@435: bool length_never_negative = false, duke@435: RegionNode* slow_region = NULL); duke@435: AllocateArrayNode* tightly_coupled_allocation(Node* ptr, duke@435: RegionNode* slow_region); duke@435: void generate_clear_array(const TypePtr* adr_type, duke@435: Node* dest, duke@435: BasicType basic_elem_type, duke@435: Node* slice_off, duke@435: Node* slice_len, duke@435: Node* slice_end); duke@435: bool generate_block_arraycopy(const TypePtr* adr_type, duke@435: BasicType basic_elem_type, duke@435: AllocateNode* alloc, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* dest_size); duke@435: void generate_slow_arraycopy(const TypePtr* adr_type, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* copy_length, duke@435: int nargs); duke@435: Node* generate_checkcast_arraycopy(const TypePtr* adr_type, duke@435: Node* dest_elem_klass, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* copy_length, int nargs); duke@435: Node* generate_generic_arraycopy(const TypePtr* adr_type, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* copy_length, int nargs); duke@435: void generate_unchecked_arraycopy(const TypePtr* adr_type, duke@435: BasicType basic_elem_type, duke@435: bool disjoint_bases, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* copy_length); duke@435: bool inline_unsafe_CAS(BasicType type); duke@435: bool inline_unsafe_ordered_store(BasicType type); duke@435: bool inline_fp_conversions(vmIntrinsics::ID id); duke@435: bool inline_reverseBytes(vmIntrinsics::ID id); duke@435: }; duke@435: duke@435: duke@435: //---------------------------make_vm_intrinsic---------------------------- duke@435: CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { duke@435: vmIntrinsics::ID id = m->intrinsic_id(); duke@435: assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); duke@435: duke@435: if (DisableIntrinsic[0] != '\0' duke@435: && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { duke@435: // disabled by a user request on the command line: duke@435: // example: -XX:DisableIntrinsic=_hashCode,_getClass duke@435: return NULL; duke@435: } duke@435: duke@435: if (!m->is_loaded()) { duke@435: // do not attempt to inline unloaded methods duke@435: return NULL; duke@435: } duke@435: duke@435: // Only a few intrinsics implement a virtual dispatch. duke@435: // They are expensive calls which are also frequently overridden. duke@435: if (is_virtual) { duke@435: switch (id) { duke@435: case vmIntrinsics::_hashCode: duke@435: case vmIntrinsics::_clone: duke@435: // OK, Object.hashCode and Object.clone intrinsics come in both flavors duke@435: break; duke@435: default: duke@435: return NULL; duke@435: } duke@435: } duke@435: duke@435: // -XX:-InlineNatives disables nearly all intrinsics: duke@435: if (!InlineNatives) { duke@435: switch (id) { duke@435: case vmIntrinsics::_indexOf: duke@435: case vmIntrinsics::_compareTo: rasbold@604: case vmIntrinsics::_equalsC: duke@435: break; // InlineNatives does not control String.compareTo duke@435: default: duke@435: return NULL; duke@435: } duke@435: } duke@435: duke@435: switch (id) { duke@435: case vmIntrinsics::_compareTo: duke@435: if (!SpecialStringCompareTo) return NULL; duke@435: break; duke@435: case vmIntrinsics::_indexOf: duke@435: if (!SpecialStringIndexOf) return NULL; duke@435: break; rasbold@604: case vmIntrinsics::_equalsC: rasbold@604: if (!SpecialArraysEquals) return NULL; rasbold@604: break; duke@435: case vmIntrinsics::_arraycopy: duke@435: if (!InlineArrayCopy) return NULL; duke@435: break; duke@435: case vmIntrinsics::_copyMemory: duke@435: if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; duke@435: if (!InlineArrayCopy) return NULL; duke@435: break; duke@435: case vmIntrinsics::_hashCode: duke@435: if (!InlineObjectHash) return NULL; duke@435: break; duke@435: case vmIntrinsics::_clone: duke@435: case vmIntrinsics::_copyOf: duke@435: case vmIntrinsics::_copyOfRange: duke@435: if (!InlineObjectCopy) return NULL; duke@435: // These also use the arraycopy intrinsic mechanism: duke@435: if (!InlineArrayCopy) return NULL; duke@435: break; duke@435: case vmIntrinsics::_checkIndex: duke@435: // We do not intrinsify this. The optimizer does fine with it. duke@435: return NULL; duke@435: duke@435: case vmIntrinsics::_get_AtomicLong: duke@435: case vmIntrinsics::_attemptUpdate: duke@435: if (!InlineAtomicLong) return NULL; duke@435: break; duke@435: duke@435: case vmIntrinsics::_Object_init: duke@435: case vmIntrinsics::_invoke: duke@435: // We do not intrinsify these; they are marked for other purposes. duke@435: return NULL; duke@435: duke@435: case vmIntrinsics::_getCallerClass: duke@435: if (!UseNewReflection) return NULL; duke@435: if (!InlineReflectionGetCallerClass) return NULL; duke@435: if (!JDK_Version::is_gte_jdk14x_version()) return NULL; duke@435: break; duke@435: duke@435: default: duke@435: break; duke@435: } duke@435: duke@435: // -XX:-InlineClassNatives disables natives from the Class class. duke@435: // The flag applies to all reflective calls, notably Array.newArray duke@435: // (visible to Java programmers as Array.newInstance). duke@435: if (m->holder()->name() == ciSymbol::java_lang_Class() || duke@435: m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { duke@435: if (!InlineClassNatives) return NULL; duke@435: } duke@435: duke@435: // -XX:-InlineThreadNatives disables natives from the Thread class. duke@435: if (m->holder()->name() == ciSymbol::java_lang_Thread()) { duke@435: if (!InlineThreadNatives) return NULL; duke@435: } duke@435: duke@435: // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. duke@435: if (m->holder()->name() == ciSymbol::java_lang_Math() || duke@435: m->holder()->name() == ciSymbol::java_lang_Float() || duke@435: m->holder()->name() == ciSymbol::java_lang_Double()) { duke@435: if (!InlineMathNatives) return NULL; duke@435: } duke@435: duke@435: // -XX:-InlineUnsafeOps disables natives from the Unsafe class. duke@435: if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { duke@435: if (!InlineUnsafeOps) return NULL; duke@435: } duke@435: duke@435: return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); duke@435: } duke@435: duke@435: //----------------------register_library_intrinsics----------------------- duke@435: // Initialize this file's data structures, for each Compile instance. duke@435: void Compile::register_library_intrinsics() { duke@435: // Nothing to do here. duke@435: } duke@435: duke@435: JVMState* LibraryIntrinsic::generate(JVMState* jvms) { duke@435: LibraryCallKit kit(jvms, this); duke@435: Compile* C = kit.C; duke@435: int nodes = C->unique(); duke@435: #ifndef PRODUCT duke@435: if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { duke@435: char buf[1000]; duke@435: const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); duke@435: tty->print_cr("Intrinsic %s", str); duke@435: } duke@435: #endif duke@435: if (kit.try_to_inline()) { duke@435: if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { duke@435: tty->print("Inlining intrinsic %s%s at bci:%d in", duke@435: vmIntrinsics::name_at(intrinsic_id()), duke@435: (is_virtual() ? " (virtual)" : ""), kit.bci()); duke@435: kit.caller()->print_short_name(tty); duke@435: tty->print_cr(" (%d bytes)", kit.caller()->code_size()); duke@435: } duke@435: C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); duke@435: if (C->log()) { duke@435: C->log()->elem("intrinsic id='%s'%s nodes='%d'", duke@435: vmIntrinsics::name_at(intrinsic_id()), duke@435: (is_virtual() ? " virtual='1'" : ""), duke@435: C->unique() - nodes); duke@435: } duke@435: return kit.transfer_exceptions_into_jvms(); duke@435: } duke@435: duke@435: if (PrintIntrinsics) { duke@435: switch (intrinsic_id()) { duke@435: case vmIntrinsics::_invoke: duke@435: case vmIntrinsics::_Object_init: duke@435: // We do not expect to inline these, so do not produce any noise about them. duke@435: break; duke@435: default: duke@435: tty->print("Did not inline intrinsic %s%s at bci:%d in", duke@435: vmIntrinsics::name_at(intrinsic_id()), duke@435: (is_virtual() ? " (virtual)" : ""), kit.bci()); duke@435: kit.caller()->print_short_name(tty); duke@435: tty->print_cr(" (%d bytes)", kit.caller()->code_size()); duke@435: } duke@435: } duke@435: C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); duke@435: return NULL; duke@435: } duke@435: duke@435: bool LibraryCallKit::try_to_inline() { duke@435: // Handle symbolic names for otherwise undistinguished boolean switches: duke@435: const bool is_store = true; duke@435: const bool is_native_ptr = true; duke@435: const bool is_static = true; duke@435: duke@435: switch (intrinsic_id()) { duke@435: case vmIntrinsics::_hashCode: duke@435: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); duke@435: case vmIntrinsics::_identityHashCode: duke@435: return inline_native_hashcode(/*!virtual*/ false, is_static); duke@435: case vmIntrinsics::_getClass: duke@435: return inline_native_getClass(); duke@435: duke@435: case vmIntrinsics::_dsin: duke@435: case vmIntrinsics::_dcos: duke@435: case vmIntrinsics::_dtan: duke@435: case vmIntrinsics::_dabs: duke@435: case vmIntrinsics::_datan2: duke@435: case vmIntrinsics::_dsqrt: duke@435: case vmIntrinsics::_dexp: duke@435: case vmIntrinsics::_dlog: duke@435: case vmIntrinsics::_dlog10: duke@435: case vmIntrinsics::_dpow: duke@435: return inline_math_native(intrinsic_id()); duke@435: duke@435: case vmIntrinsics::_min: duke@435: case vmIntrinsics::_max: duke@435: return inline_min_max(intrinsic_id()); duke@435: duke@435: case vmIntrinsics::_arraycopy: duke@435: return inline_arraycopy(); duke@435: duke@435: case vmIntrinsics::_compareTo: duke@435: return inline_string_compareTo(); duke@435: case vmIntrinsics::_indexOf: duke@435: return inline_string_indexOf(); duke@435: duke@435: case vmIntrinsics::_getObject: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); duke@435: case vmIntrinsics::_getBoolean: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); duke@435: case vmIntrinsics::_getByte: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); duke@435: case vmIntrinsics::_getShort: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); duke@435: case vmIntrinsics::_getChar: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); duke@435: case vmIntrinsics::_getInt: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); duke@435: case vmIntrinsics::_getLong: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); duke@435: case vmIntrinsics::_getFloat: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); duke@435: case vmIntrinsics::_getDouble: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); duke@435: duke@435: case vmIntrinsics::_putObject: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); duke@435: case vmIntrinsics::_putBoolean: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); duke@435: case vmIntrinsics::_putByte: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); duke@435: case vmIntrinsics::_putShort: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); duke@435: case vmIntrinsics::_putChar: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); duke@435: case vmIntrinsics::_putInt: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); duke@435: case vmIntrinsics::_putLong: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); duke@435: case vmIntrinsics::_putFloat: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); duke@435: case vmIntrinsics::_putDouble: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); duke@435: duke@435: case vmIntrinsics::_getByte_raw: duke@435: return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); duke@435: case vmIntrinsics::_getShort_raw: duke@435: return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); duke@435: case vmIntrinsics::_getChar_raw: duke@435: return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); duke@435: case vmIntrinsics::_getInt_raw: duke@435: return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); duke@435: case vmIntrinsics::_getLong_raw: duke@435: return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); duke@435: case vmIntrinsics::_getFloat_raw: duke@435: return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); duke@435: case vmIntrinsics::_getDouble_raw: duke@435: return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); duke@435: case vmIntrinsics::_getAddress_raw: duke@435: return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); duke@435: duke@435: case vmIntrinsics::_putByte_raw: duke@435: return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); duke@435: case vmIntrinsics::_putShort_raw: duke@435: return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); duke@435: case vmIntrinsics::_putChar_raw: duke@435: return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); duke@435: case vmIntrinsics::_putInt_raw: duke@435: return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); duke@435: case vmIntrinsics::_putLong_raw: duke@435: return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); duke@435: case vmIntrinsics::_putFloat_raw: duke@435: return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); duke@435: case vmIntrinsics::_putDouble_raw: duke@435: return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); duke@435: case vmIntrinsics::_putAddress_raw: duke@435: return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); duke@435: duke@435: case vmIntrinsics::_getObjectVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); duke@435: case vmIntrinsics::_getBooleanVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); duke@435: case vmIntrinsics::_getByteVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); duke@435: case vmIntrinsics::_getShortVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); duke@435: case vmIntrinsics::_getCharVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); duke@435: case vmIntrinsics::_getIntVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); duke@435: case vmIntrinsics::_getLongVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); duke@435: case vmIntrinsics::_getFloatVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); duke@435: case vmIntrinsics::_getDoubleVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); duke@435: duke@435: case vmIntrinsics::_putObjectVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); duke@435: case vmIntrinsics::_putBooleanVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); duke@435: case vmIntrinsics::_putByteVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); duke@435: case vmIntrinsics::_putShortVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); duke@435: case vmIntrinsics::_putCharVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); duke@435: case vmIntrinsics::_putIntVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); duke@435: case vmIntrinsics::_putLongVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); duke@435: case vmIntrinsics::_putFloatVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); duke@435: case vmIntrinsics::_putDoubleVolatile: duke@435: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); duke@435: duke@435: case vmIntrinsics::_prefetchRead: duke@435: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); duke@435: case vmIntrinsics::_prefetchWrite: duke@435: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); duke@435: case vmIntrinsics::_prefetchReadStatic: duke@435: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); duke@435: case vmIntrinsics::_prefetchWriteStatic: duke@435: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); duke@435: duke@435: case vmIntrinsics::_compareAndSwapObject: duke@435: return inline_unsafe_CAS(T_OBJECT); duke@435: case vmIntrinsics::_compareAndSwapInt: duke@435: return inline_unsafe_CAS(T_INT); duke@435: case vmIntrinsics::_compareAndSwapLong: duke@435: return inline_unsafe_CAS(T_LONG); duke@435: duke@435: case vmIntrinsics::_putOrderedObject: duke@435: return inline_unsafe_ordered_store(T_OBJECT); duke@435: case vmIntrinsics::_putOrderedInt: duke@435: return inline_unsafe_ordered_store(T_INT); duke@435: case vmIntrinsics::_putOrderedLong: duke@435: return inline_unsafe_ordered_store(T_LONG); duke@435: duke@435: case vmIntrinsics::_currentThread: duke@435: return inline_native_currentThread(); duke@435: case vmIntrinsics::_isInterrupted: duke@435: return inline_native_isInterrupted(); duke@435: duke@435: case vmIntrinsics::_currentTimeMillis: duke@435: return inline_native_time_funcs(false); duke@435: case vmIntrinsics::_nanoTime: duke@435: return inline_native_time_funcs(true); duke@435: case vmIntrinsics::_allocateInstance: duke@435: return inline_unsafe_allocate(); duke@435: case vmIntrinsics::_copyMemory: duke@435: return inline_unsafe_copyMemory(); duke@435: case vmIntrinsics::_newArray: duke@435: return inline_native_newArray(); duke@435: case vmIntrinsics::_getLength: duke@435: return inline_native_getLength(); duke@435: case vmIntrinsics::_copyOf: duke@435: return inline_array_copyOf(false); duke@435: case vmIntrinsics::_copyOfRange: duke@435: return inline_array_copyOf(true); rasbold@604: case vmIntrinsics::_equalsC: rasbold@604: return inline_array_equals(); duke@435: case vmIntrinsics::_clone: duke@435: return inline_native_clone(intrinsic()->is_virtual()); duke@435: duke@435: case vmIntrinsics::_isAssignableFrom: duke@435: return inline_native_subtype_check(); duke@435: duke@435: case vmIntrinsics::_isInstance: duke@435: case vmIntrinsics::_getModifiers: duke@435: case vmIntrinsics::_isInterface: duke@435: case vmIntrinsics::_isArray: duke@435: case vmIntrinsics::_isPrimitive: duke@435: case vmIntrinsics::_getSuperclass: duke@435: case vmIntrinsics::_getComponentType: duke@435: case vmIntrinsics::_getClassAccessFlags: duke@435: return inline_native_Class_query(intrinsic_id()); duke@435: duke@435: case vmIntrinsics::_floatToRawIntBits: duke@435: case vmIntrinsics::_floatToIntBits: duke@435: case vmIntrinsics::_intBitsToFloat: duke@435: case vmIntrinsics::_doubleToRawLongBits: duke@435: case vmIntrinsics::_doubleToLongBits: duke@435: case vmIntrinsics::_longBitsToDouble: duke@435: return inline_fp_conversions(intrinsic_id()); duke@435: duke@435: case vmIntrinsics::_reverseBytes_i: duke@435: case vmIntrinsics::_reverseBytes_l: duke@435: return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); duke@435: duke@435: case vmIntrinsics::_get_AtomicLong: duke@435: return inline_native_AtomicLong_get(); duke@435: case vmIntrinsics::_attemptUpdate: duke@435: return inline_native_AtomicLong_attemptUpdate(); duke@435: duke@435: case vmIntrinsics::_getCallerClass: duke@435: return inline_native_Reflection_getCallerClass(); duke@435: duke@435: default: duke@435: // If you get here, it may be that someone has added a new intrinsic duke@435: // to the list in vmSymbols.hpp without implementing it here. duke@435: #ifndef PRODUCT duke@435: if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { duke@435: tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", duke@435: vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); duke@435: } duke@435: #endif duke@435: return false; duke@435: } duke@435: } duke@435: duke@435: //------------------------------push_result------------------------------ duke@435: // Helper function for finishing intrinsics. duke@435: void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { duke@435: record_for_igvn(region); duke@435: set_control(_gvn.transform(region)); duke@435: BasicType value_type = value->type()->basic_type(); duke@435: push_node(value_type, _gvn.transform(value)); duke@435: } duke@435: duke@435: //------------------------------generate_guard--------------------------- duke@435: // Helper function for generating guarded fast-slow graph structures. duke@435: // The given 'test', if true, guards a slow path. If the test fails duke@435: // then a fast path can be taken. (We generally hope it fails.) duke@435: // In all cases, GraphKit::control() is updated to the fast path. duke@435: // The returned value represents the control for the slow path. duke@435: // The return value is never 'top'; it is either a valid control duke@435: // or NULL if it is obvious that the slow path can never be taken. duke@435: // Also, if region and the slow control are not NULL, the slow edge duke@435: // is appended to the region. duke@435: Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { duke@435: if (stopped()) { duke@435: // Already short circuited. duke@435: return NULL; duke@435: } duke@435: duke@435: // Build an if node and its projections. duke@435: // If test is true we take the slow path, which we assume is uncommon. duke@435: if (_gvn.type(test) == TypeInt::ZERO) { duke@435: // The slow branch is never taken. No need to build this guard. duke@435: return NULL; duke@435: } duke@435: duke@435: IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); duke@435: duke@435: Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) ); duke@435: if (if_slow == top()) { duke@435: // The slow branch is never taken. No need to build this guard. duke@435: return NULL; duke@435: } duke@435: duke@435: if (region != NULL) duke@435: region->add_req(if_slow); duke@435: duke@435: Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) ); duke@435: set_control(if_fast); duke@435: duke@435: return if_slow; duke@435: } duke@435: duke@435: inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { duke@435: return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); duke@435: } duke@435: inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { duke@435: return generate_guard(test, region, PROB_FAIR); duke@435: } duke@435: duke@435: inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, duke@435: Node* *pos_index) { duke@435: if (stopped()) duke@435: return NULL; // already stopped duke@435: if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] duke@435: return NULL; // index is already adequately typed duke@435: Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); duke@435: Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); duke@435: Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); duke@435: if (is_neg != NULL && pos_index != NULL) { duke@435: // Emulate effect of Parse::adjust_map_after_if. duke@435: Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS); duke@435: ccast->set_req(0, control()); duke@435: (*pos_index) = _gvn.transform(ccast); duke@435: } duke@435: return is_neg; duke@435: } duke@435: duke@435: inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, duke@435: Node* *pos_index) { duke@435: if (stopped()) duke@435: return NULL; // already stopped duke@435: if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] duke@435: return NULL; // index is already adequately typed duke@435: Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); duke@435: BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); duke@435: Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) ); duke@435: Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); duke@435: if (is_notp != NULL && pos_index != NULL) { duke@435: // Emulate effect of Parse::adjust_map_after_if. duke@435: Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1); duke@435: ccast->set_req(0, control()); duke@435: (*pos_index) = _gvn.transform(ccast); duke@435: } duke@435: return is_notp; duke@435: } duke@435: duke@435: // Make sure that 'position' is a valid limit index, in [0..length]. duke@435: // There are two equivalent plans for checking this: duke@435: // A. (offset + copyLength) unsigned<= arrayLength duke@435: // B. offset <= (arrayLength - copyLength) duke@435: // We require that all of the values above, except for the sum and duke@435: // difference, are already known to be non-negative. duke@435: // Plan A is robust in the face of overflow, if offset and copyLength duke@435: // are both hugely positive. duke@435: // duke@435: // Plan B is less direct and intuitive, but it does not overflow at duke@435: // all, since the difference of two non-negatives is always duke@435: // representable. Whenever Java methods must perform the equivalent duke@435: // check they generally use Plan B instead of Plan A. duke@435: // For the moment we use Plan A. duke@435: inline Node* LibraryCallKit::generate_limit_guard(Node* offset, duke@435: Node* subseq_length, duke@435: Node* array_length, duke@435: RegionNode* region) { duke@435: if (stopped()) duke@435: return NULL; // already stopped duke@435: bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; duke@435: if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length)) duke@435: return NULL; // common case of whole-array copy duke@435: Node* last = subseq_length; duke@435: if (!zero_offset) // last += offset duke@435: last = _gvn.transform( new (C, 3) AddINode(last, offset)); duke@435: Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) ); duke@435: Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); duke@435: Node* is_over = generate_guard(bol_lt, region, PROB_MIN); duke@435: return is_over; duke@435: } duke@435: duke@435: duke@435: //--------------------------generate_current_thread-------------------- duke@435: Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { duke@435: ciKlass* thread_klass = env()->Thread_klass(); duke@435: const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); duke@435: Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode()); duke@435: Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); duke@435: Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); duke@435: tls_output = thread; duke@435: return threadObj; duke@435: } duke@435: duke@435: duke@435: //------------------------------inline_string_compareTo------------------------ duke@435: bool LibraryCallKit::inline_string_compareTo() { duke@435: duke@435: const int value_offset = java_lang_String::value_offset_in_bytes(); duke@435: const int count_offset = java_lang_String::count_offset_in_bytes(); duke@435: const int offset_offset = java_lang_String::offset_offset_in_bytes(); duke@435: duke@435: _sp += 2; duke@435: Node *argument = pop(); // pop non-receiver first: it was pushed second duke@435: Node *receiver = pop(); duke@435: duke@435: // Null check on self without removing any arguments. The argument duke@435: // null check technically happens in the wrong place, which can lead to duke@435: // invalid stack traces when string compare is inlined into a method duke@435: // which handles NullPointerExceptions. duke@435: _sp += 2; duke@435: receiver = do_null_check(receiver, T_OBJECT); duke@435: argument = do_null_check(argument, T_OBJECT); duke@435: _sp -= 2; duke@435: if (stopped()) { duke@435: return true; duke@435: } duke@435: duke@435: ciInstanceKlass* klass = env()->String_klass(); duke@435: const TypeInstPtr* string_type = duke@435: TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); duke@435: duke@435: Node* compare = duke@435: _gvn.transform(new (C, 7) StrCompNode( duke@435: control(), duke@435: memory(TypeAryPtr::CHARS), duke@435: memory(string_type->add_offset(value_offset)), duke@435: memory(string_type->add_offset(count_offset)), duke@435: memory(string_type->add_offset(offset_offset)), duke@435: receiver, duke@435: argument)); duke@435: push(compare); duke@435: return true; duke@435: } duke@435: rasbold@604: //------------------------------inline_array_equals---------------------------- rasbold@604: bool LibraryCallKit::inline_array_equals() { rasbold@604: rasbold@609: if (!Matcher::has_match_rule(Op_AryEq)) return false; rasbold@609: rasbold@604: _sp += 2; rasbold@604: Node *argument2 = pop(); rasbold@604: Node *argument1 = pop(); rasbold@604: rasbold@604: Node* equals = rasbold@604: _gvn.transform(new (C, 3) AryEqNode(control(), rasbold@604: argument1, rasbold@604: argument2) rasbold@604: ); rasbold@604: push(equals); rasbold@604: return true; rasbold@604: } rasbold@604: duke@435: // Java version of String.indexOf(constant string) duke@435: // class StringDecl { duke@435: // StringDecl(char[] ca) { duke@435: // offset = 0; duke@435: // count = ca.length; duke@435: // value = ca; duke@435: // } duke@435: // int offset; duke@435: // int count; duke@435: // char[] value; duke@435: // } duke@435: // duke@435: // static int string_indexOf_J(StringDecl string_object, char[] target_object, duke@435: // int targetOffset, int cache_i, int md2) { duke@435: // int cache = cache_i; duke@435: // int sourceOffset = string_object.offset; duke@435: // int sourceCount = string_object.count; duke@435: // int targetCount = target_object.length; duke@435: // duke@435: // int targetCountLess1 = targetCount - 1; duke@435: // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; duke@435: // duke@435: // char[] source = string_object.value; duke@435: // char[] target = target_object; duke@435: // int lastChar = target[targetCountLess1]; duke@435: // duke@435: // outer_loop: duke@435: // for (int i = sourceOffset; i < sourceEnd; ) { duke@435: // int src = source[i + targetCountLess1]; duke@435: // if (src == lastChar) { duke@435: // // With random strings and a 4-character alphabet, duke@435: // // reverse matching at this point sets up 0.8% fewer duke@435: // // frames, but (paradoxically) makes 0.3% more probes. duke@435: // // Since those probes are nearer the lastChar probe, duke@435: // // there is may be a net D$ win with reverse matching. duke@435: // // But, reversing loop inhibits unroll of inner loop duke@435: // // for unknown reason. So, does running outer loop from duke@435: // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) duke@435: // for (int j = 0; j < targetCountLess1; j++) { duke@435: // if (target[targetOffset + j] != source[i+j]) { duke@435: // if ((cache & (1 << source[i+j])) == 0) { duke@435: // if (md2 < j+1) { duke@435: // i += j+1; duke@435: // continue outer_loop; duke@435: // } duke@435: // } duke@435: // i += md2; duke@435: // continue outer_loop; duke@435: // } duke@435: // } duke@435: // return i - sourceOffset; duke@435: // } duke@435: // if ((cache & (1 << src)) == 0) { duke@435: // i += targetCountLess1; duke@435: // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. duke@435: // i++; duke@435: // } duke@435: // return -1; duke@435: // } duke@435: duke@435: //------------------------------string_indexOf------------------------ duke@435: Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, duke@435: jint cache_i, jint md2_i) { duke@435: duke@435: Node* no_ctrl = NULL; duke@435: float likely = PROB_LIKELY(0.9); duke@435: float unlikely = PROB_UNLIKELY(0.9); duke@435: duke@435: const int value_offset = java_lang_String::value_offset_in_bytes(); duke@435: const int count_offset = java_lang_String::count_offset_in_bytes(); duke@435: const int offset_offset = java_lang_String::offset_offset_in_bytes(); duke@435: duke@435: ciInstanceKlass* klass = env()->String_klass(); duke@435: const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); duke@435: const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0); duke@435: duke@435: Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset); duke@435: Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); duke@435: Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset); duke@435: Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); duke@435: Node* sourcea = basic_plus_adr(string_object, string_object, value_offset); duke@435: Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset)); duke@435: kvn@599: Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) ); duke@435: jint target_length = target_array->length(); duke@435: const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); duke@435: const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); duke@435: duke@435: IdealKit kit(gvn(), control(), merged_memory()); duke@435: #define __ kit. duke@435: Node* zero = __ ConI(0); duke@435: Node* one = __ ConI(1); duke@435: Node* cache = __ ConI(cache_i); duke@435: Node* md2 = __ ConI(md2_i); duke@435: Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); duke@435: Node* targetCount = __ ConI(target_length); duke@435: Node* targetCountLess1 = __ ConI(target_length - 1); duke@435: Node* targetOffset = __ ConI(targetOffset_i); duke@435: Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); duke@435: duke@435: IdealVariable rtn(kit), i(kit), j(kit); __ declares_done(); duke@435: Node* outer_loop = __ make_label(2 /* goto */); duke@435: Node* return_ = __ make_label(1); duke@435: duke@435: __ set(rtn,__ ConI(-1)); duke@435: __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); { duke@435: Node* i2 = __ AddI(__ value(i), targetCountLess1); duke@435: // pin to prohibit loading of "next iteration" value which may SEGV (rare) duke@435: Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); duke@435: __ if_then(src, BoolTest::eq, lastChar, unlikely); { duke@435: __ loop(j, zero, BoolTest::lt, targetCountLess1); { duke@435: Node* tpj = __ AddI(targetOffset, __ value(j)); duke@435: Node* targ = load_array_element(no_ctrl, target, tpj, target_type); duke@435: Node* ipj = __ AddI(__ value(i), __ value(j)); duke@435: Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); duke@435: __ if_then(targ, BoolTest::ne, src2); { duke@435: __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { duke@435: __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { duke@435: __ increment(i, __ AddI(__ value(j), one)); duke@435: __ goto_(outer_loop); duke@435: } __ end_if(); __ dead(j); duke@435: }__ end_if(); __ dead(j); duke@435: __ increment(i, md2); duke@435: __ goto_(outer_loop); duke@435: }__ end_if(); duke@435: __ increment(j, one); duke@435: }__ end_loop(); __ dead(j); duke@435: __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); duke@435: __ goto_(return_); duke@435: }__ end_if(); duke@435: __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { duke@435: __ increment(i, targetCountLess1); duke@435: }__ end_if(); duke@435: __ increment(i, one); duke@435: __ bind(outer_loop); duke@435: }__ end_loop(); __ dead(i); duke@435: __ bind(return_); duke@435: __ drain_delay_transform(); duke@435: duke@435: set_control(__ ctrl()); duke@435: Node* result = __ value(rtn); duke@435: #undef __ duke@435: C->set_has_loops(true); duke@435: return result; duke@435: } duke@435: duke@435: duke@435: //------------------------------inline_string_indexOf------------------------ duke@435: bool LibraryCallKit::inline_string_indexOf() { duke@435: duke@435: _sp += 2; duke@435: Node *argument = pop(); // pop non-receiver first: it was pushed second duke@435: Node *receiver = pop(); duke@435: twisti@1040: // don't intrinsify if argument isn't a constant string. duke@435: if (!argument->is_Con()) { duke@435: return false; duke@435: } duke@435: const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); duke@435: if (str_type == NULL) { duke@435: return false; duke@435: } duke@435: ciInstanceKlass* klass = env()->String_klass(); duke@435: ciObject* str_const = str_type->const_oop(); duke@435: if (str_const == NULL || str_const->klass() != klass) { duke@435: return false; duke@435: } duke@435: ciInstance* str = str_const->as_instance(); duke@435: assert(str != NULL, "must be instance"); duke@435: duke@435: const int value_offset = java_lang_String::value_offset_in_bytes(); duke@435: const int count_offset = java_lang_String::count_offset_in_bytes(); duke@435: const int offset_offset = java_lang_String::offset_offset_in_bytes(); duke@435: duke@435: ciObject* v = str->field_value_by_offset(value_offset).as_object(); duke@435: int o = str->field_value_by_offset(offset_offset).as_int(); duke@435: int c = str->field_value_by_offset(count_offset).as_int(); duke@435: ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array duke@435: duke@435: // constant strings have no offset and count == length which duke@435: // simplifies the resulting code somewhat so lets optimize for that. duke@435: if (o != 0 || c != pat->length()) { duke@435: return false; duke@435: } duke@435: duke@435: // Null check on self without removing any arguments. The argument duke@435: // null check technically happens in the wrong place, which can lead to duke@435: // invalid stack traces when string compare is inlined into a method duke@435: // which handles NullPointerExceptions. duke@435: _sp += 2; duke@435: receiver = do_null_check(receiver, T_OBJECT); duke@435: // No null check on the argument is needed since it's a constant String oop. duke@435: _sp -= 2; duke@435: if (stopped()) { duke@435: return true; duke@435: } duke@435: duke@435: // The null string as a pattern always returns 0 (match at beginning of string) duke@435: if (c == 0) { duke@435: push(intcon(0)); duke@435: return true; duke@435: } duke@435: duke@435: jchar lastChar = pat->char_at(o + (c - 1)); duke@435: int cache = 0; duke@435: int i; duke@435: for (i = 0; i < c - 1; i++) { duke@435: assert(i < pat->length(), "out of range"); duke@435: cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); duke@435: } duke@435: duke@435: int md2 = c; duke@435: for (i = 0; i < c - 1; i++) { duke@435: assert(i < pat->length(), "out of range"); duke@435: if (pat->char_at(o + i) == lastChar) { duke@435: md2 = (c - 1) - i; duke@435: } duke@435: } duke@435: duke@435: Node* result = string_indexOf(receiver, pat, o, cache, md2); duke@435: push(result); duke@435: return true; duke@435: } duke@435: duke@435: //--------------------------pop_math_arg-------------------------------- duke@435: // Pop a double argument to a math function from the stack duke@435: // rounding it if necessary. duke@435: Node * LibraryCallKit::pop_math_arg() { duke@435: Node *arg = pop_pair(); duke@435: if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) duke@435: arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) ); duke@435: return arg; duke@435: } duke@435: duke@435: //------------------------------inline_trig---------------------------------- duke@435: // Inline sin/cos/tan instructions, if possible. If rounding is required, do duke@435: // argument reduction which will turn into a fast/slow diamond. duke@435: bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { duke@435: _sp += arg_size(); // restore stack pointer duke@435: Node* arg = pop_math_arg(); duke@435: Node* trig = NULL; duke@435: duke@435: switch (id) { duke@435: case vmIntrinsics::_dsin: duke@435: trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg)); duke@435: break; duke@435: case vmIntrinsics::_dcos: duke@435: trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg)); duke@435: break; duke@435: case vmIntrinsics::_dtan: duke@435: trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg)); duke@435: break; duke@435: default: duke@435: assert(false, "bad intrinsic was passed in"); duke@435: return false; duke@435: } duke@435: duke@435: // Rounding required? Check for argument reduction! duke@435: if( Matcher::strict_fp_requires_explicit_rounding ) { duke@435: duke@435: static const double pi_4 = 0.7853981633974483; duke@435: static const double neg_pi_4 = -0.7853981633974483; duke@435: // pi/2 in 80-bit extended precision duke@435: // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; duke@435: // -pi/2 in 80-bit extended precision duke@435: // 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}; duke@435: // Cutoff value for using this argument reduction technique duke@435: //static const double pi_2_minus_epsilon = 1.564660403643354; duke@435: //static const double neg_pi_2_plus_epsilon = -1.564660403643354; duke@435: duke@435: // Pseudocode for sin: duke@435: // if (x <= Math.PI / 4.0) { duke@435: // if (x >= -Math.PI / 4.0) return fsin(x); duke@435: // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); duke@435: // } else { duke@435: // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); duke@435: // } duke@435: // return StrictMath.sin(x); duke@435: duke@435: // Pseudocode for cos: duke@435: // if (x <= Math.PI / 4.0) { duke@435: // if (x >= -Math.PI / 4.0) return fcos(x); duke@435: // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); duke@435: // } else { duke@435: // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); duke@435: // } duke@435: // return StrictMath.cos(x); duke@435: duke@435: // Actually, sticking in an 80-bit Intel value into C2 will be tough; it duke@435: // requires a special machine instruction to load it. Instead we'll try duke@435: // the 'easy' case. If we really need the extra range +/- PI/2 we'll duke@435: // probably do the math inside the SIN encoding. duke@435: duke@435: // Make the merge point duke@435: RegionNode *r = new (C, 3) RegionNode(3); duke@435: Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE); duke@435: duke@435: // Flatten arg so we need only 1 test duke@435: Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg)); duke@435: // Node for PI/4 constant duke@435: Node *pi4 = makecon(TypeD::make(pi_4)); duke@435: // Check PI/4 : abs(arg) duke@435: Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs)); duke@435: // Check: If PI/4 < abs(arg) then go slow duke@435: Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) ); duke@435: // Branch either way duke@435: IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); duke@435: set_control(opt_iff(r,iff)); duke@435: duke@435: // Set fast path result duke@435: phi->init_req(2,trig); duke@435: duke@435: // Slow path - non-blocking leaf call duke@435: Node* call = NULL; duke@435: switch (id) { duke@435: case vmIntrinsics::_dsin: duke@435: call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), duke@435: CAST_FROM_FN_PTR(address, SharedRuntime::dsin), duke@435: "Sin", NULL, arg, top()); duke@435: break; duke@435: case vmIntrinsics::_dcos: duke@435: call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), duke@435: CAST_FROM_FN_PTR(address, SharedRuntime::dcos), duke@435: "Cos", NULL, arg, top()); duke@435: break; duke@435: case vmIntrinsics::_dtan: duke@435: call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), duke@435: CAST_FROM_FN_PTR(address, SharedRuntime::dtan), duke@435: "Tan", NULL, arg, top()); duke@435: break; duke@435: } duke@435: assert(control()->in(0) == call, ""); duke@435: Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms)); duke@435: r->init_req(1,control()); duke@435: phi->init_req(1,slow_result); duke@435: duke@435: // Post-merge duke@435: set_control(_gvn.transform(r)); duke@435: record_for_igvn(r); duke@435: trig = _gvn.transform(phi); duke@435: duke@435: C->set_has_split_ifs(true); // Has chance for split-if optimization duke@435: } duke@435: // Push result back on JVM stack duke@435: push_pair(trig); duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------inline_sqrt------------------------------------- duke@435: // Inline square root instruction, if possible. duke@435: bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { duke@435: assert(id == vmIntrinsics::_dsqrt, "Not square root"); duke@435: _sp += arg_size(); // restore stack pointer duke@435: push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg()))); duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------inline_abs------------------------------------- duke@435: // Inline absolute value instruction, if possible. duke@435: bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { duke@435: assert(id == vmIntrinsics::_dabs, "Not absolute value"); duke@435: _sp += arg_size(); // restore stack pointer duke@435: push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg()))); duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------inline_exp------------------------------------- duke@435: // Inline exp instructions, if possible. The Intel hardware only misses duke@435: // really odd corner cases (+/- Infinity). Just uncommon-trap them. duke@435: bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { duke@435: assert(id == vmIntrinsics::_dexp, "Not exp"); duke@435: duke@435: // If this inlining ever returned NaN in the past, we do not intrinsify it duke@435: // every again. NaN results requires StrictMath.exp handling. duke@435: if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; duke@435: duke@435: // Do not intrinsify on older platforms which lack cmove. duke@435: if (ConditionalMoveLimit == 0) return false; duke@435: duke@435: _sp += arg_size(); // restore stack pointer duke@435: Node *x = pop_math_arg(); duke@435: Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x)); duke@435: duke@435: //------------------- duke@435: //result=(result.isNaN())? StrictMath::exp():result; duke@435: // Check: If isNaN() by checking result!=result? then go to Strict Math duke@435: Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); duke@435: // Build the boolean node duke@435: Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); duke@435: duke@435: { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); duke@435: // End the current control-flow path duke@435: push_pair(x); duke@435: // Math.exp intrinsic returned a NaN, which requires StrictMath.exp duke@435: // to handle. Recompile without intrinsifying Math.exp duke@435: uncommon_trap(Deoptimization::Reason_intrinsic, duke@435: Deoptimization::Action_make_not_entrant); duke@435: } duke@435: duke@435: C->set_has_split_ifs(true); // Has chance for split-if optimization duke@435: duke@435: push_pair(result); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------inline_pow------------------------------------- duke@435: // Inline power instructions, if possible. duke@435: bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { duke@435: assert(id == vmIntrinsics::_dpow, "Not pow"); duke@435: duke@435: // If this inlining ever returned NaN in the past, we do not intrinsify it duke@435: // every again. NaN results requires StrictMath.pow handling. duke@435: if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; duke@435: duke@435: // Do not intrinsify on older platforms which lack cmove. duke@435: if (ConditionalMoveLimit == 0) return false; duke@435: duke@435: // Pseudocode for pow duke@435: // if (x <= 0.0) { duke@435: // if ((double)((int)y)==y) { // if y is int duke@435: // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y) duke@435: // } else { duke@435: // result = NaN; duke@435: // } duke@435: // } else { duke@435: // result = DPow(x,y); duke@435: // } duke@435: // if (result != result)? { twisti@1040: // uncommon_trap(); duke@435: // } duke@435: // return result; duke@435: duke@435: _sp += arg_size(); // restore stack pointer duke@435: Node* y = pop_math_arg(); duke@435: Node* x = pop_math_arg(); duke@435: duke@435: Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); duke@435: duke@435: // Short form: if not top-level (i.e., Math.pow but inlining Math.pow duke@435: // inside of something) then skip the fancy tests and just check for duke@435: // NaN result. duke@435: Node *result = NULL; duke@435: if( jvms()->depth() >= 1 ) { duke@435: result = fast_result; duke@435: } else { duke@435: duke@435: // Set the merge point for If node with condition of (x <= 0.0) duke@435: // There are four possible paths to region node and phi node duke@435: RegionNode *r = new (C, 4) RegionNode(4); duke@435: Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE); duke@435: duke@435: // Build the first if node: if (x <= 0.0) duke@435: // Node for 0 constant duke@435: Node *zeronode = makecon(TypeD::ZERO); duke@435: // Check x:0 duke@435: Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode)); duke@435: // Check: If (x<=0) then go complex path duke@435: Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) ); duke@435: // Branch either way duke@435: IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); duke@435: Node *opt_test = _gvn.transform(if1); duke@435: //assert( opt_test->is_If(), "Expect an IfNode"); duke@435: IfNode *opt_if1 = (IfNode*)opt_test; duke@435: // Fast path taken; set region slot 3 duke@435: Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) ); duke@435: r->init_req(3,fast_taken); // Capture fast-control duke@435: duke@435: // Fast path not-taken, i.e. slow path duke@435: Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) ); duke@435: duke@435: // Set fast path result duke@435: Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) ); duke@435: phi->init_req(3, fast_result); duke@435: duke@435: // Complex path duke@435: // Build the second if node (if y is int) duke@435: // Node for (int)y duke@435: Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y)); duke@435: // Node for (double)((int) y) duke@435: Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty)); duke@435: // Check (double)((int) y) : y duke@435: Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y)); duke@435: // Check if (y isn't int) then go to slow path duke@435: duke@435: Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) ); twisti@1040: // Branch either way duke@435: IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); duke@435: Node *slow_path = opt_iff(r,if2); // Set region path 2 duke@435: duke@435: // Calculate DPow(abs(x), y)*(1 & (int)y) duke@435: // Node for constant 1 duke@435: Node *conone = intcon(1); duke@435: // 1& (int)y duke@435: Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) ); duke@435: // zero node duke@435: Node *conzero = intcon(0); duke@435: // Check (1&(int)y)==0? duke@435: Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero)); duke@435: // Check if (1&(int)y)!=0?, if so the result is negative duke@435: Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) ); duke@435: // abs(x) duke@435: Node *absx=_gvn.transform( new (C, 2) AbsDNode(x)); duke@435: // abs(x)^y duke@435: Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) ); duke@435: // -abs(x)^y duke@435: Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy)); duke@435: // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y) duke@435: Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); duke@435: // Set complex path fast result duke@435: phi->init_req(2, signresult); duke@435: duke@435: static const jlong nan_bits = CONST64(0x7ff8000000000000); duke@435: Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN duke@435: r->init_req(1,slow_path); duke@435: phi->init_req(1,slow_result); duke@435: duke@435: // Post merge duke@435: set_control(_gvn.transform(r)); duke@435: record_for_igvn(r); duke@435: result=_gvn.transform(phi); duke@435: } duke@435: duke@435: //------------------- duke@435: //result=(result.isNaN())? uncommon_trap():result; duke@435: // Check: If isNaN() by checking result!=result? then go to Strict Math duke@435: Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); duke@435: // Build the boolean node duke@435: Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); duke@435: duke@435: { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); duke@435: // End the current control-flow path duke@435: push_pair(x); duke@435: push_pair(y); duke@435: // Math.pow intrinsic returned a NaN, which requires StrictMath.pow duke@435: // to handle. Recompile without intrinsifying Math.pow. duke@435: uncommon_trap(Deoptimization::Reason_intrinsic, duke@435: Deoptimization::Action_make_not_entrant); duke@435: } duke@435: duke@435: C->set_has_split_ifs(true); // Has chance for split-if optimization duke@435: duke@435: push_pair(result); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------inline_trans------------------------------------- duke@435: // Inline transcendental instructions, if possible. The Intel hardware gets duke@435: // these right, no funny corner cases missed. duke@435: bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { duke@435: _sp += arg_size(); // restore stack pointer duke@435: Node* arg = pop_math_arg(); duke@435: Node* trans = NULL; duke@435: duke@435: switch (id) { duke@435: case vmIntrinsics::_dlog: duke@435: trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg)); duke@435: break; duke@435: case vmIntrinsics::_dlog10: duke@435: trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg)); duke@435: break; duke@435: default: duke@435: assert(false, "bad intrinsic was passed in"); duke@435: return false; duke@435: } duke@435: duke@435: // Push result back on JVM stack duke@435: push_pair(trans); duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------runtime_math----------------------------- duke@435: bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { duke@435: Node* a = NULL; duke@435: Node* b = NULL; duke@435: duke@435: assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), duke@435: "must be (DD)D or (D)D type"); duke@435: duke@435: // Inputs duke@435: _sp += arg_size(); // restore stack pointer duke@435: if (call_type == OptoRuntime::Math_DD_D_Type()) { duke@435: b = pop_math_arg(); duke@435: } duke@435: a = pop_math_arg(); duke@435: duke@435: const TypePtr* no_memory_effects = NULL; duke@435: Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, duke@435: no_memory_effects, duke@435: a, top(), b, b ? top() : NULL); duke@435: Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0)); duke@435: #ifdef ASSERT duke@435: Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1)); duke@435: assert(value_top == top(), "second value must be top"); duke@435: #endif duke@435: duke@435: push_pair(value); duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------inline_math_native----------------------------- duke@435: bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { duke@435: switch (id) { duke@435: // These intrinsics are not properly supported on all hardware duke@435: case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : duke@435: runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); duke@435: case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : duke@435: runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); duke@435: case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : duke@435: runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); duke@435: duke@435: case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : duke@435: runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); duke@435: case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : duke@435: runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); duke@435: duke@435: // These intrinsics are supported on all hardware duke@435: case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; duke@435: case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; duke@435: duke@435: // These intrinsics don't work on X86. The ad implementation doesn't duke@435: // handle NaN's properly. Instead of returning infinity, the ad duke@435: // implementation returns a NaN on overflow. See bug: 6304089 duke@435: // Once the ad implementations are fixed, change the code below duke@435: // to match the intrinsics above duke@435: duke@435: case vmIntrinsics::_dexp: return duke@435: runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); duke@435: case vmIntrinsics::_dpow: return duke@435: runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); duke@435: duke@435: // These intrinsics are not yet correctly implemented duke@435: case vmIntrinsics::_datan2: duke@435: return false; duke@435: duke@435: default: duke@435: ShouldNotReachHere(); duke@435: return false; duke@435: } duke@435: } duke@435: duke@435: static bool is_simple_name(Node* n) { duke@435: return (n->req() == 1 // constant duke@435: || (n->is_Type() && n->as_Type()->type()->singleton()) duke@435: || n->is_Proj() // parameter or return value duke@435: || n->is_Phi() // local of some sort duke@435: ); duke@435: } duke@435: duke@435: //----------------------------inline_min_max----------------------------------- duke@435: bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { duke@435: push(generate_min_max(id, argument(0), argument(1))); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: Node* duke@435: LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { duke@435: // These are the candidate return value: duke@435: Node* xvalue = x0; duke@435: Node* yvalue = y0; duke@435: duke@435: if (xvalue == yvalue) { duke@435: return xvalue; duke@435: } duke@435: duke@435: bool want_max = (id == vmIntrinsics::_max); duke@435: duke@435: const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); duke@435: const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); duke@435: if (txvalue == NULL || tyvalue == NULL) return top(); duke@435: // This is not really necessary, but it is consistent with a duke@435: // hypothetical MaxINode::Value method: duke@435: int widen = MAX2(txvalue->_widen, tyvalue->_widen); duke@435: duke@435: // %%% This folding logic should (ideally) be in a different place. duke@435: // Some should be inside IfNode, and there to be a more reliable duke@435: // transformation of ?: style patterns into cmoves. We also want duke@435: // more powerful optimizations around cmove and min/max. duke@435: duke@435: // Try to find a dominating comparison of these guys. duke@435: // It can simplify the index computation for Arrays.copyOf duke@435: // and similar uses of System.arraycopy. duke@435: // First, compute the normalized version of CmpI(x, y). duke@435: int cmp_op = Op_CmpI; duke@435: Node* xkey = xvalue; duke@435: Node* ykey = yvalue; duke@435: Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) ); duke@435: if (ideal_cmpxy->is_Cmp()) { duke@435: // E.g., if we have CmpI(length - offset, count), duke@435: // it might idealize to CmpI(length, count + offset) duke@435: cmp_op = ideal_cmpxy->Opcode(); duke@435: xkey = ideal_cmpxy->in(1); duke@435: ykey = ideal_cmpxy->in(2); duke@435: } duke@435: duke@435: // Start by locating any relevant comparisons. duke@435: Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; duke@435: Node* cmpxy = NULL; duke@435: Node* cmpyx = NULL; duke@435: for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { duke@435: Node* cmp = start_from->fast_out(k); duke@435: if (cmp->outcnt() > 0 && // must have prior uses duke@435: cmp->in(0) == NULL && // must be context-independent duke@435: cmp->Opcode() == cmp_op) { // right kind of compare duke@435: if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; duke@435: if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; duke@435: } duke@435: } duke@435: duke@435: const int NCMPS = 2; duke@435: Node* cmps[NCMPS] = { cmpxy, cmpyx }; duke@435: int cmpn; duke@435: for (cmpn = 0; cmpn < NCMPS; cmpn++) { duke@435: if (cmps[cmpn] != NULL) break; // find a result duke@435: } duke@435: if (cmpn < NCMPS) { duke@435: // Look for a dominating test that tells us the min and max. duke@435: int depth = 0; // Limit search depth for speed duke@435: Node* dom = control(); duke@435: for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { duke@435: if (++depth >= 100) break; duke@435: Node* ifproj = dom; duke@435: if (!ifproj->is_Proj()) continue; duke@435: Node* iff = ifproj->in(0); duke@435: if (!iff->is_If()) continue; duke@435: Node* bol = iff->in(1); duke@435: if (!bol->is_Bool()) continue; duke@435: Node* cmp = bol->in(1); duke@435: if (cmp == NULL) continue; duke@435: for (cmpn = 0; cmpn < NCMPS; cmpn++) duke@435: if (cmps[cmpn] == cmp) break; duke@435: if (cmpn == NCMPS) continue; duke@435: BoolTest::mask btest = bol->as_Bool()->_test._test; duke@435: if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); duke@435: if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); duke@435: // At this point, we know that 'x btest y' is true. duke@435: switch (btest) { duke@435: case BoolTest::eq: duke@435: // They are proven equal, so we can collapse the min/max. duke@435: // Either value is the answer. Choose the simpler. duke@435: if (is_simple_name(yvalue) && !is_simple_name(xvalue)) duke@435: return yvalue; duke@435: return xvalue; duke@435: case BoolTest::lt: // x < y duke@435: case BoolTest::le: // x <= y duke@435: return (want_max ? yvalue : xvalue); duke@435: case BoolTest::gt: // x > y duke@435: case BoolTest::ge: // x >= y duke@435: return (want_max ? xvalue : yvalue); duke@435: } duke@435: } duke@435: } duke@435: duke@435: // We failed to find a dominating test. duke@435: // Let's pick a test that might GVN with prior tests. duke@435: Node* best_bol = NULL; duke@435: BoolTest::mask best_btest = BoolTest::illegal; duke@435: for (cmpn = 0; cmpn < NCMPS; cmpn++) { duke@435: Node* cmp = cmps[cmpn]; duke@435: if (cmp == NULL) continue; duke@435: for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { duke@435: Node* bol = cmp->fast_out(j); duke@435: if (!bol->is_Bool()) continue; duke@435: BoolTest::mask btest = bol->as_Bool()->_test._test; duke@435: if (btest == BoolTest::eq || btest == BoolTest::ne) continue; duke@435: if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); duke@435: if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { duke@435: best_bol = bol->as_Bool(); duke@435: best_btest = btest; duke@435: } duke@435: } duke@435: } duke@435: duke@435: Node* answer_if_true = NULL; duke@435: Node* answer_if_false = NULL; duke@435: switch (best_btest) { duke@435: default: duke@435: if (cmpxy == NULL) duke@435: cmpxy = ideal_cmpxy; duke@435: best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) ); duke@435: // and fall through: duke@435: case BoolTest::lt: // x < y duke@435: case BoolTest::le: // x <= y duke@435: answer_if_true = (want_max ? yvalue : xvalue); duke@435: answer_if_false = (want_max ? xvalue : yvalue); duke@435: break; duke@435: case BoolTest::gt: // x > y duke@435: case BoolTest::ge: // x >= y duke@435: answer_if_true = (want_max ? xvalue : yvalue); duke@435: answer_if_false = (want_max ? yvalue : xvalue); duke@435: break; duke@435: } duke@435: duke@435: jint hi, lo; duke@435: if (want_max) { duke@435: // We can sharpen the minimum. duke@435: hi = MAX2(txvalue->_hi, tyvalue->_hi); duke@435: lo = MAX2(txvalue->_lo, tyvalue->_lo); duke@435: } else { duke@435: // We can sharpen the maximum. duke@435: hi = MIN2(txvalue->_hi, tyvalue->_hi); duke@435: lo = MIN2(txvalue->_lo, tyvalue->_lo); duke@435: } duke@435: duke@435: // Use a flow-free graph structure, to avoid creating excess control edges duke@435: // which could hinder other optimizations. duke@435: // Since Math.min/max is often used with arraycopy, we want duke@435: // tightly_coupled_allocation to be able to see beyond min/max expressions. duke@435: Node* cmov = CMoveNode::make(C, NULL, best_bol, duke@435: answer_if_false, answer_if_true, duke@435: TypeInt::make(lo, hi, widen)); duke@435: duke@435: return _gvn.transform(cmov); duke@435: duke@435: /* duke@435: // This is not as desirable as it may seem, since Min and Max duke@435: // nodes do not have a full set of optimizations. duke@435: // And they would interfere, anyway, with 'if' optimizations duke@435: // and with CMoveI canonical forms. duke@435: switch (id) { duke@435: case vmIntrinsics::_min: duke@435: result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; duke@435: case vmIntrinsics::_max: duke@435: result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; duke@435: default: duke@435: ShouldNotReachHere(); duke@435: } duke@435: */ duke@435: } duke@435: duke@435: inline int duke@435: LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { duke@435: const TypePtr* base_type = TypePtr::NULL_PTR; duke@435: if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); duke@435: if (base_type == NULL) { duke@435: // Unknown type. duke@435: return Type::AnyPtr; duke@435: } else if (base_type == TypePtr::NULL_PTR) { duke@435: // Since this is a NULL+long form, we have to switch to a rawptr. duke@435: base = _gvn.transform( new (C, 2) CastX2PNode(offset) ); duke@435: offset = MakeConX(0); duke@435: return Type::RawPtr; duke@435: } else if (base_type->base() == Type::RawPtr) { duke@435: return Type::RawPtr; duke@435: } else if (base_type->isa_oopptr()) { duke@435: // Base is never null => always a heap address. duke@435: if (base_type->ptr() == TypePtr::NotNull) { duke@435: return Type::OopPtr; duke@435: } duke@435: // Offset is small => always a heap address. duke@435: const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); duke@435: if (offset_type != NULL && duke@435: base_type->offset() == 0 && // (should always be?) duke@435: offset_type->_lo >= 0 && duke@435: !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { duke@435: return Type::OopPtr; duke@435: } duke@435: // Otherwise, it might either be oop+off or NULL+addr. duke@435: return Type::AnyPtr; duke@435: } else { duke@435: // No information: duke@435: return Type::AnyPtr; duke@435: } duke@435: } duke@435: duke@435: inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { duke@435: int kind = classify_unsafe_addr(base, offset); duke@435: if (kind == Type::RawPtr) { duke@435: return basic_plus_adr(top(), base, offset); duke@435: } else { duke@435: return basic_plus_adr(base, offset); duke@435: } duke@435: } duke@435: duke@435: //----------------------------inline_reverseBytes_int/long------------------- twisti@1040: // inline Integer.reverseBytes(int) twisti@1040: // inline Long.reverseBytes(long) duke@435: bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { duke@435: assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes"); duke@435: if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; duke@435: if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; duke@435: _sp += arg_size(); // restore stack pointer duke@435: switch (id) { duke@435: case vmIntrinsics::_reverseBytes_i: duke@435: push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop()))); duke@435: break; duke@435: case vmIntrinsics::_reverseBytes_l: duke@435: push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair()))); duke@435: break; duke@435: default: duke@435: ; duke@435: } duke@435: return true; duke@435: } duke@435: duke@435: //----------------------------inline_unsafe_access---------------------------- duke@435: duke@435: const static BasicType T_ADDRESS_HOLDER = T_LONG; duke@435: duke@435: // Interpret Unsafe.fieldOffset cookies correctly: duke@435: extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); duke@435: duke@435: bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { duke@435: if (callee()->is_static()) return false; // caller must have the capability! duke@435: duke@435: #ifndef PRODUCT duke@435: { duke@435: ResourceMark rm; duke@435: // Check the signatures. duke@435: ciSignature* sig = signature(); duke@435: #ifdef ASSERT duke@435: if (!is_store) { duke@435: // Object getObject(Object base, int/long offset), etc. duke@435: BasicType rtype = sig->return_type()->basic_type(); duke@435: if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) duke@435: rtype = T_ADDRESS; // it is really a C void* duke@435: assert(rtype == type, "getter must return the expected value"); duke@435: if (!is_native_ptr) { duke@435: assert(sig->count() == 2, "oop getter has 2 arguments"); duke@435: assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); duke@435: assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); duke@435: } else { duke@435: assert(sig->count() == 1, "native getter has 1 argument"); duke@435: assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); duke@435: } duke@435: } else { duke@435: // void putObject(Object base, int/long offset, Object x), etc. duke@435: assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); duke@435: if (!is_native_ptr) { duke@435: assert(sig->count() == 3, "oop putter has 3 arguments"); duke@435: assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); duke@435: assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); duke@435: } else { duke@435: assert(sig->count() == 2, "native putter has 2 arguments"); duke@435: assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); duke@435: } duke@435: BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); duke@435: if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) duke@435: vtype = T_ADDRESS; // it is really a C void* duke@435: assert(vtype == type, "putter must accept the expected value"); duke@435: } duke@435: #endif // ASSERT duke@435: } duke@435: #endif //PRODUCT duke@435: duke@435: C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". duke@435: duke@435: int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; duke@435: duke@435: // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words duke@435: int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); duke@435: duke@435: debug_only(int saved_sp = _sp); duke@435: _sp += nargs; duke@435: duke@435: Node* val; duke@435: debug_only(val = (Node*)(uintptr_t)-1); duke@435: duke@435: duke@435: if (is_store) { duke@435: // Get the value being stored. (Pop it first; it was pushed last.) duke@435: switch (type) { duke@435: case T_DOUBLE: duke@435: case T_LONG: duke@435: case T_ADDRESS: duke@435: val = pop_pair(); duke@435: break; duke@435: default: duke@435: val = pop(); duke@435: } duke@435: } duke@435: duke@435: // Build address expression. See the code in inline_unsafe_prefetch. duke@435: Node *adr; duke@435: Node *heap_base_oop = top(); duke@435: if (!is_native_ptr) { duke@435: // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset duke@435: Node* offset = pop_pair(); duke@435: // The base is either a Java object or a value produced by Unsafe.staticFieldBase duke@435: Node* base = pop(); duke@435: // We currently rely on the cookies produced by Unsafe.xxxFieldOffset duke@435: // to be plain byte offsets, which are also the same as those accepted duke@435: // by oopDesc::field_base. duke@435: assert(Unsafe_field_offset_to_byte_offset(11) == 11, duke@435: "fieldOffset must be byte-scaled"); duke@435: // 32-bit machines ignore the high half! duke@435: offset = ConvL2X(offset); duke@435: adr = make_unsafe_address(base, offset); duke@435: heap_base_oop = base; duke@435: } else { duke@435: Node* ptr = pop_pair(); duke@435: // Adjust Java long to machine word: duke@435: ptr = ConvL2X(ptr); duke@435: adr = make_unsafe_address(NULL, ptr); duke@435: } duke@435: duke@435: // Pop receiver last: it was pushed first. duke@435: Node *receiver = pop(); duke@435: duke@435: assert(saved_sp == _sp, "must have correct argument count"); duke@435: duke@435: const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); duke@435: duke@435: // First guess at the value type. duke@435: const Type *value_type = Type::get_const_basic_type(type); duke@435: duke@435: // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, duke@435: // there was not enough information to nail it down. duke@435: Compile::AliasType* alias_type = C->alias_type(adr_type); duke@435: assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); duke@435: duke@435: // We will need memory barriers unless we can determine a unique duke@435: // alias category for this reference. (Note: If for some reason duke@435: // the barriers get omitted and the unsafe reference begins to "pollute" duke@435: // the alias analysis of the rest of the graph, either Compile::can_alias duke@435: // or Compile::must_alias will throw a diagnostic assert.) duke@435: bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); duke@435: duke@435: if (!is_store && type == T_OBJECT) { duke@435: // Attempt to infer a sharper value type from the offset and base type. duke@435: ciKlass* sharpened_klass = NULL; duke@435: duke@435: // See if it is an instance field, with an object type. duke@435: if (alias_type->field() != NULL) { duke@435: assert(!is_native_ptr, "native pointer op cannot use a java address"); duke@435: if (alias_type->field()->type()->is_klass()) { duke@435: sharpened_klass = alias_type->field()->type()->as_klass(); duke@435: } duke@435: } duke@435: duke@435: // See if it is a narrow oop array. duke@435: if (adr_type->isa_aryptr()) { coleenp@548: if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes(type)) { duke@435: const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); duke@435: if (elem_type != NULL) { duke@435: sharpened_klass = elem_type->klass(); duke@435: } duke@435: } duke@435: } duke@435: duke@435: if (sharpened_klass != NULL) { duke@435: const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); duke@435: duke@435: // Sharpen the value type. duke@435: value_type = tjp; duke@435: duke@435: #ifndef PRODUCT duke@435: if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { duke@435: tty->print(" from base type: "); adr_type->dump(); duke@435: tty->print(" sharpened value: "); value_type->dump(); duke@435: } duke@435: #endif duke@435: } duke@435: } duke@435: duke@435: // Null check on self without removing any arguments. The argument duke@435: // null check technically happens in the wrong place, which can lead to duke@435: // invalid stack traces when the primitive is inlined into a method duke@435: // which handles NullPointerExceptions. duke@435: _sp += nargs; duke@435: do_null_check(receiver, T_OBJECT); duke@435: _sp -= nargs; duke@435: if (stopped()) { duke@435: return true; duke@435: } duke@435: // Heap pointers get a null-check from the interpreter, duke@435: // as a courtesy. However, this is not guaranteed by Unsafe, duke@435: // and it is not possible to fully distinguish unintended nulls duke@435: // from intended ones in this API. duke@435: duke@435: if (is_volatile) { duke@435: // We need to emit leading and trailing CPU membars (see below) in duke@435: // addition to memory membars when is_volatile. This is a little duke@435: // too strong, but avoids the need to insert per-alias-type duke@435: // volatile membars (for stores; compare Parse::do_put_xxx), which twisti@1040: // we cannot do effectively here because we probably only have a duke@435: // rough approximation of type. duke@435: need_mem_bar = true; duke@435: // For Stores, place a memory ordering barrier now. duke@435: if (is_store) duke@435: insert_mem_bar(Op_MemBarRelease); duke@435: } duke@435: duke@435: // Memory barrier to prevent normal and 'unsafe' accesses from duke@435: // bypassing each other. Happens after null checks, so the duke@435: // exception paths do not take memory state from the memory barrier, duke@435: // so there's no problems making a strong assert about mixing users duke@435: // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar duke@435: // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. duke@435: if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); duke@435: duke@435: if (!is_store) { duke@435: Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); duke@435: // load value and push onto stack duke@435: switch (type) { duke@435: case T_BOOLEAN: duke@435: case T_CHAR: duke@435: case T_BYTE: duke@435: case T_SHORT: duke@435: case T_INT: duke@435: case T_FLOAT: duke@435: case T_OBJECT: duke@435: push( p ); duke@435: break; duke@435: case T_ADDRESS: duke@435: // Cast to an int type. duke@435: p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) ); duke@435: p = ConvX2L(p); duke@435: push_pair(p); duke@435: break; duke@435: case T_DOUBLE: duke@435: case T_LONG: duke@435: push_pair( p ); duke@435: break; duke@435: default: ShouldNotReachHere(); duke@435: } duke@435: } else { duke@435: // place effect of store into memory duke@435: switch (type) { duke@435: case T_DOUBLE: duke@435: val = dstore_rounding(val); duke@435: break; duke@435: case T_ADDRESS: duke@435: // Repackage the long as a pointer. duke@435: val = ConvL2X(val); duke@435: val = _gvn.transform( new (C, 2) CastX2PNode(val) ); duke@435: break; duke@435: } duke@435: duke@435: if (type != T_OBJECT ) { duke@435: (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); duke@435: } else { duke@435: // Possibly an oop being stored to Java heap or native memory duke@435: if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { duke@435: // oop to Java heap. duke@435: (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type); duke@435: } else { duke@435: duke@435: // We can't tell at compile time if we are storing in the Java heap or outside duke@435: // of it. So we need to emit code to conditionally do the proper type of duke@435: // store. duke@435: duke@435: IdealKit kit(gvn(), control(), merged_memory()); duke@435: kit.declares_done(); duke@435: // QQQ who knows what probability is here?? duke@435: kit.if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { duke@435: (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type); duke@435: } kit.else_(); { duke@435: (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); duke@435: } kit.end_if(); duke@435: } duke@435: } duke@435: } duke@435: duke@435: if (is_volatile) { duke@435: if (!is_store) duke@435: insert_mem_bar(Op_MemBarAcquire); duke@435: else duke@435: insert_mem_bar(Op_MemBarVolatile); duke@435: } duke@435: duke@435: if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //----------------------------inline_unsafe_prefetch---------------------------- duke@435: duke@435: bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { duke@435: #ifndef PRODUCT duke@435: { duke@435: ResourceMark rm; duke@435: // Check the signatures. duke@435: ciSignature* sig = signature(); duke@435: #ifdef ASSERT duke@435: // Object getObject(Object base, int/long offset), etc. duke@435: BasicType rtype = sig->return_type()->basic_type(); duke@435: if (!is_native_ptr) { duke@435: assert(sig->count() == 2, "oop prefetch has 2 arguments"); duke@435: assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); duke@435: assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); duke@435: } else { duke@435: assert(sig->count() == 1, "native prefetch has 1 argument"); duke@435: assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); duke@435: } duke@435: #endif // ASSERT duke@435: } duke@435: #endif // !PRODUCT duke@435: duke@435: C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". duke@435: duke@435: // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args duke@435: int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); duke@435: duke@435: debug_only(int saved_sp = _sp); duke@435: _sp += nargs; duke@435: duke@435: // Build address expression. See the code in inline_unsafe_access. duke@435: Node *adr; duke@435: if (!is_native_ptr) { duke@435: // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset duke@435: Node* offset = pop_pair(); duke@435: // The base is either a Java object or a value produced by Unsafe.staticFieldBase duke@435: Node* base = pop(); duke@435: // We currently rely on the cookies produced by Unsafe.xxxFieldOffset duke@435: // to be plain byte offsets, which are also the same as those accepted duke@435: // by oopDesc::field_base. duke@435: assert(Unsafe_field_offset_to_byte_offset(11) == 11, duke@435: "fieldOffset must be byte-scaled"); duke@435: // 32-bit machines ignore the high half! duke@435: offset = ConvL2X(offset); duke@435: adr = make_unsafe_address(base, offset); duke@435: } else { duke@435: Node* ptr = pop_pair(); duke@435: // Adjust Java long to machine word: duke@435: ptr = ConvL2X(ptr); duke@435: adr = make_unsafe_address(NULL, ptr); duke@435: } duke@435: duke@435: if (is_static) { duke@435: assert(saved_sp == _sp, "must have correct argument count"); duke@435: } else { duke@435: // Pop receiver last: it was pushed first. duke@435: Node *receiver = pop(); duke@435: assert(saved_sp == _sp, "must have correct argument count"); duke@435: duke@435: // Null check on self without removing any arguments. The argument duke@435: // null check technically happens in the wrong place, which can lead to duke@435: // invalid stack traces when the primitive is inlined into a method duke@435: // which handles NullPointerExceptions. duke@435: _sp += nargs; duke@435: do_null_check(receiver, T_OBJECT); duke@435: _sp -= nargs; duke@435: if (stopped()) { duke@435: return true; duke@435: } duke@435: } duke@435: duke@435: // Generate the read or write prefetch duke@435: Node *prefetch; duke@435: if (is_store) { duke@435: prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr); duke@435: } else { duke@435: prefetch = new (C, 3) PrefetchReadNode(i_o(), adr); duke@435: } duke@435: prefetch->init_req(0, control()); duke@435: set_i_o(_gvn.transform(prefetch)); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //----------------------------inline_unsafe_CAS---------------------------- duke@435: duke@435: bool LibraryCallKit::inline_unsafe_CAS(BasicType type) { duke@435: // This basic scheme here is the same as inline_unsafe_access, but duke@435: // differs in enough details that combining them would make the code duke@435: // overly confusing. (This is a true fact! I originally combined duke@435: // them, but even I was confused by it!) As much code/comments as duke@435: // possible are retained from inline_unsafe_access though to make twisti@1040: // the correspondences clearer. - dl duke@435: duke@435: if (callee()->is_static()) return false; // caller must have the capability! duke@435: duke@435: #ifndef PRODUCT duke@435: { duke@435: ResourceMark rm; duke@435: // Check the signatures. duke@435: ciSignature* sig = signature(); duke@435: #ifdef ASSERT duke@435: BasicType rtype = sig->return_type()->basic_type(); duke@435: assert(rtype == T_BOOLEAN, "CAS must return boolean"); duke@435: assert(sig->count() == 4, "CAS has 4 arguments"); duke@435: assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); duke@435: assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); duke@435: #endif // ASSERT duke@435: } duke@435: #endif //PRODUCT duke@435: duke@435: // number of stack slots per value argument (1 or 2) duke@435: int type_words = type2size[type]; duke@435: duke@435: // Cannot inline wide CAS on machines that don't support it natively kvn@464: if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8()) duke@435: return false; duke@435: duke@435: C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". duke@435: duke@435: // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue; duke@435: int nargs = 1 + 1 + 2 + type_words + type_words; duke@435: duke@435: // pop arguments: newval, oldval, offset, base, and receiver duke@435: debug_only(int saved_sp = _sp); duke@435: _sp += nargs; duke@435: Node* newval = (type_words == 1) ? pop() : pop_pair(); duke@435: Node* oldval = (type_words == 1) ? pop() : pop_pair(); duke@435: Node *offset = pop_pair(); duke@435: Node *base = pop(); duke@435: Node *receiver = pop(); duke@435: assert(saved_sp == _sp, "must have correct argument count"); duke@435: duke@435: // Null check receiver. duke@435: _sp += nargs; duke@435: do_null_check(receiver, T_OBJECT); duke@435: _sp -= nargs; duke@435: if (stopped()) { duke@435: return true; duke@435: } duke@435: duke@435: // Build field offset expression. duke@435: // We currently rely on the cookies produced by Unsafe.xxxFieldOffset duke@435: // to be plain byte offsets, which are also the same as those accepted duke@435: // by oopDesc::field_base. duke@435: assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); duke@435: // 32-bit machines ignore the high half of long offsets duke@435: offset = ConvL2X(offset); duke@435: Node* adr = make_unsafe_address(base, offset); duke@435: const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); duke@435: duke@435: // (Unlike inline_unsafe_access, there seems no point in trying duke@435: // to refine types. Just use the coarse types here. duke@435: const Type *value_type = Type::get_const_basic_type(type); duke@435: Compile::AliasType* alias_type = C->alias_type(adr_type); duke@435: assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); duke@435: int alias_idx = C->get_alias_index(adr_type); duke@435: duke@435: // Memory-model-wise, a CAS acts like a little synchronized block, twisti@1040: // so needs barriers on each side. These don't translate into duke@435: // actual barriers on most machines, but we still need rest of duke@435: // compiler to respect ordering. duke@435: duke@435: insert_mem_bar(Op_MemBarRelease); duke@435: insert_mem_bar(Op_MemBarCPUOrder); duke@435: duke@435: // 4984716: MemBars must be inserted before this duke@435: // memory node in order to avoid a false duke@435: // dependency which will confuse the scheduler. duke@435: Node *mem = memory(alias_idx); duke@435: duke@435: // For now, we handle only those cases that actually exist: ints, duke@435: // longs, and Object. Adding others should be straightforward. duke@435: Node* cas; duke@435: switch(type) { duke@435: case T_INT: duke@435: cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval)); duke@435: break; duke@435: case T_LONG: duke@435: cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); duke@435: break; duke@435: case T_OBJECT: coleenp@548: // reference stores need a store barrier. duke@435: // (They don't if CAS fails, but it isn't worth checking.) duke@435: pre_barrier(control(), base, adr, alias_idx, newval, value_type, T_OBJECT); coleenp@548: #ifdef _LP64 kvn@598: if (adr->bottom_type()->is_ptr_to_narrowoop()) { kvn@656: Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); kvn@656: Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); coleenp@548: cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr, kvn@656: newval_enc, oldval_enc)); coleenp@548: } else coleenp@548: #endif kvn@656: { kvn@656: cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); kvn@656: } duke@435: post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true); duke@435: break; duke@435: default: duke@435: ShouldNotReachHere(); duke@435: break; duke@435: } duke@435: duke@435: // SCMemProjNodes represent the memory state of CAS. Their main duke@435: // role is to prevent CAS nodes from being optimized away when their duke@435: // results aren't used. duke@435: Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); duke@435: set_memory(proj, alias_idx); duke@435: duke@435: // Add the trailing membar surrounding the access duke@435: insert_mem_bar(Op_MemBarCPUOrder); duke@435: insert_mem_bar(Op_MemBarAcquire); duke@435: duke@435: push(cas); duke@435: return true; duke@435: } duke@435: duke@435: bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { duke@435: // This is another variant of inline_unsafe_access, differing in duke@435: // that it always issues store-store ("release") barrier and ensures duke@435: // store-atomicity (which only matters for "long"). duke@435: duke@435: if (callee()->is_static()) return false; // caller must have the capability! duke@435: duke@435: #ifndef PRODUCT duke@435: { duke@435: ResourceMark rm; duke@435: // Check the signatures. duke@435: ciSignature* sig = signature(); duke@435: #ifdef ASSERT duke@435: BasicType rtype = sig->return_type()->basic_type(); duke@435: assert(rtype == T_VOID, "must return void"); duke@435: assert(sig->count() == 3, "has 3 arguments"); duke@435: assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); duke@435: assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); duke@435: #endif // ASSERT duke@435: } duke@435: #endif //PRODUCT duke@435: duke@435: // number of stack slots per value argument (1 or 2) duke@435: int type_words = type2size[type]; duke@435: duke@435: C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". duke@435: duke@435: // Argument words: "this" plus oop plus offset plus value; duke@435: int nargs = 1 + 1 + 2 + type_words; duke@435: duke@435: // pop arguments: val, offset, base, and receiver duke@435: debug_only(int saved_sp = _sp); duke@435: _sp += nargs; duke@435: Node* val = (type_words == 1) ? pop() : pop_pair(); duke@435: Node *offset = pop_pair(); duke@435: Node *base = pop(); duke@435: Node *receiver = pop(); duke@435: assert(saved_sp == _sp, "must have correct argument count"); duke@435: duke@435: // Null check receiver. duke@435: _sp += nargs; duke@435: do_null_check(receiver, T_OBJECT); duke@435: _sp -= nargs; duke@435: if (stopped()) { duke@435: return true; duke@435: } duke@435: duke@435: // Build field offset expression. duke@435: assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); duke@435: // 32-bit machines ignore the high half of long offsets duke@435: offset = ConvL2X(offset); duke@435: Node* adr = make_unsafe_address(base, offset); duke@435: const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); duke@435: const Type *value_type = Type::get_const_basic_type(type); duke@435: Compile::AliasType* alias_type = C->alias_type(adr_type); duke@435: duke@435: insert_mem_bar(Op_MemBarRelease); duke@435: insert_mem_bar(Op_MemBarCPUOrder); duke@435: // Ensure that the store is atomic for longs: duke@435: bool require_atomic_access = true; duke@435: Node* store; duke@435: if (type == T_OBJECT) // reference stores need a store barrier. duke@435: store = store_oop_to_unknown(control(), base, adr, adr_type, val, value_type, type); duke@435: else { duke@435: store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); duke@435: } duke@435: insert_mem_bar(Op_MemBarCPUOrder); duke@435: return true; duke@435: } duke@435: duke@435: bool LibraryCallKit::inline_unsafe_allocate() { duke@435: if (callee()->is_static()) return false; // caller must have the capability! duke@435: int nargs = 1 + 1; duke@435: assert(signature()->size() == nargs-1, "alloc has 1 argument"); duke@435: null_check_receiver(callee()); // check then ignore argument(0) duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: Node* cls = do_null_check(argument(1), T_OBJECT); duke@435: _sp -= nargs; duke@435: if (stopped()) return true; duke@435: duke@435: Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: kls = do_null_check(kls, T_OBJECT); duke@435: _sp -= nargs; duke@435: if (stopped()) return true; // argument was like int.class duke@435: duke@435: // Note: The argument might still be an illegal value like duke@435: // Serializable.class or Object[].class. The runtime will handle it. duke@435: // But we must make an explicit check for initialization. duke@435: Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc)); duke@435: Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT); duke@435: Node* bits = intcon(instanceKlass::fully_initialized); duke@435: Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) ); duke@435: // The 'test' is non-zero if we need to take a slow path. duke@435: duke@435: Node* obj = new_instance(kls, test); duke@435: push(obj); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //------------------------inline_native_time_funcs-------------- duke@435: // inline code for System.currentTimeMillis() and System.nanoTime() duke@435: // these have the same type and signature duke@435: bool LibraryCallKit::inline_native_time_funcs(bool isNano) { duke@435: address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) : duke@435: CAST_FROM_FN_PTR(address, os::javaTimeMillis); duke@435: const char * funcName = isNano ? "nanoTime" : "currentTimeMillis"; duke@435: const TypeFunc *tf = OptoRuntime::current_time_millis_Type(); duke@435: const TypePtr* no_memory_effects = NULL; duke@435: Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); duke@435: Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0)); duke@435: #ifdef ASSERT duke@435: Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1)); duke@435: assert(value_top == top(), "second value must be top"); duke@435: #endif duke@435: push_pair(value); duke@435: return true; duke@435: } duke@435: duke@435: //------------------------inline_native_currentThread------------------ duke@435: bool LibraryCallKit::inline_native_currentThread() { duke@435: Node* junk = NULL; duke@435: push(generate_current_thread(junk)); duke@435: return true; duke@435: } duke@435: duke@435: //------------------------inline_native_isInterrupted------------------ duke@435: bool LibraryCallKit::inline_native_isInterrupted() { duke@435: const int nargs = 1+1; // receiver + boolean duke@435: assert(nargs == arg_size(), "sanity"); duke@435: // Add a fast path to t.isInterrupted(clear_int): duke@435: // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) duke@435: // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) duke@435: // So, in the common case that the interrupt bit is false, duke@435: // we avoid making a call into the VM. Even if the interrupt bit duke@435: // is true, if the clear_int argument is false, we avoid the VM call. duke@435: // However, if the receiver is not currentThread, we must call the VM, duke@435: // because there must be some locking done around the operation. duke@435: duke@435: // We only go to the fast case code if we pass two guards. duke@435: // Paths which do not pass are accumulated in the slow_region. duke@435: RegionNode* slow_region = new (C, 1) RegionNode(1); duke@435: record_for_igvn(slow_region); duke@435: RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow duke@435: PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL); duke@435: enum { no_int_result_path = 1, duke@435: no_clear_result_path = 2, duke@435: slow_result_path = 3 duke@435: }; duke@435: duke@435: // (a) Receiving thread must be the current thread. duke@435: Node* rec_thr = argument(0); duke@435: Node* tls_ptr = NULL; duke@435: Node* cur_thr = generate_current_thread(tls_ptr); duke@435: Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) ); duke@435: Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) ); duke@435: duke@435: bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); duke@435: if (!known_current_thread) duke@435: generate_slow_guard(bol_thr, slow_region); duke@435: duke@435: // (b) Interrupt bit on TLS must be false. duke@435: Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); duke@435: Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); duke@435: p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); duke@435: Node* int_bit = make_load(NULL, p, TypeInt::BOOL, T_INT); duke@435: Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) ); duke@435: Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) ); duke@435: duke@435: IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); duke@435: duke@435: // First fast path: if (!TLS._interrupted) return false; duke@435: Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) ); duke@435: result_rgn->init_req(no_int_result_path, false_bit); duke@435: result_val->init_req(no_int_result_path, intcon(0)); duke@435: duke@435: // drop through to next case duke@435: set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) ); duke@435: duke@435: // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. duke@435: Node* clr_arg = argument(1); duke@435: Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) ); duke@435: Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) ); duke@435: IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); duke@435: duke@435: // Second fast path: ... else if (!clear_int) return true; duke@435: Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) ); duke@435: result_rgn->init_req(no_clear_result_path, false_arg); duke@435: result_val->init_req(no_clear_result_path, intcon(1)); duke@435: duke@435: // drop through to next case duke@435: set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) ); duke@435: duke@435: // (d) Otherwise, go to the slow path. duke@435: slow_region->add_req(control()); duke@435: set_control( _gvn.transform(slow_region) ); duke@435: duke@435: if (stopped()) { duke@435: // There is no slow path. duke@435: result_rgn->init_req(slow_result_path, top()); duke@435: result_val->init_req(slow_result_path, top()); duke@435: } else { duke@435: // non-virtual because it is a private non-static duke@435: CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); duke@435: duke@435: Node* slow_val = set_results_for_java_call(slow_call); duke@435: // this->control() comes from set_results_for_java_call duke@435: duke@435: // If we know that the result of the slow call will be true, tell the optimizer! duke@435: if (known_current_thread) slow_val = intcon(1); duke@435: duke@435: Node* fast_io = slow_call->in(TypeFunc::I_O); duke@435: Node* fast_mem = slow_call->in(TypeFunc::Memory); duke@435: // These two phis are pre-filled with copies of of the fast IO and Memory duke@435: Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); duke@435: Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); duke@435: duke@435: result_rgn->init_req(slow_result_path, control()); duke@435: io_phi ->init_req(slow_result_path, i_o()); duke@435: mem_phi ->init_req(slow_result_path, reset_memory()); duke@435: result_val->init_req(slow_result_path, slow_val); duke@435: duke@435: set_all_memory( _gvn.transform(mem_phi) ); duke@435: set_i_o( _gvn.transform(io_phi) ); duke@435: } duke@435: duke@435: push_result(result_rgn, result_val); duke@435: C->set_has_split_ifs(true); // Has chance for split-if optimization duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //---------------------------load_mirror_from_klass---------------------------- duke@435: // Given a klass oop, load its java mirror (a java.lang.Class oop). duke@435: Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { duke@435: Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc)); duke@435: return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); duke@435: } duke@435: duke@435: //-----------------------load_klass_from_mirror_common------------------------- duke@435: // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. duke@435: // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), duke@435: // and branch to the given path on the region. duke@435: // If never_see_null, take an uncommon trap on null, so we can optimistically duke@435: // compile for the non-null case. duke@435: // If the region is NULL, force never_see_null = true. duke@435: Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, duke@435: bool never_see_null, duke@435: int nargs, duke@435: RegionNode* region, duke@435: int null_path, duke@435: int offset) { duke@435: if (region == NULL) never_see_null = true; duke@435: Node* p = basic_plus_adr(mirror, offset); duke@435: const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; kvn@599: Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) ); duke@435: _sp += nargs; // any deopt will start just before call to enclosing method duke@435: Node* null_ctl = top(); duke@435: kls = null_check_oop(kls, &null_ctl, never_see_null); duke@435: if (region != NULL) { duke@435: // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). duke@435: region->init_req(null_path, null_ctl); duke@435: } else { duke@435: assert(null_ctl == top(), "no loose ends"); duke@435: } duke@435: _sp -= nargs; duke@435: return kls; duke@435: } duke@435: duke@435: //--------------------(inline_native_Class_query helpers)--------------------- duke@435: // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. duke@435: // Fall through if (mods & mask) == bits, take the guard otherwise. duke@435: Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { duke@435: // Branch around if the given klass has the given modifier bit set. duke@435: // Like generate_guard, adds a new path onto the region. duke@435: Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); duke@435: Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); duke@435: Node* mask = intcon(modifier_mask); duke@435: Node* bits = intcon(modifier_bits); duke@435: Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) ); duke@435: Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) ); duke@435: Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) ); duke@435: return generate_fair_guard(bol, region); duke@435: } duke@435: Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { duke@435: return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); duke@435: } duke@435: duke@435: //-------------------------inline_native_Class_query------------------- duke@435: bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { duke@435: int nargs = 1+0; // just the Class mirror, in most cases duke@435: const Type* return_type = TypeInt::BOOL; duke@435: Node* prim_return_value = top(); // what happens if it's a primitive class? duke@435: bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); duke@435: bool expect_prim = false; // most of these guys expect to work on refs duke@435: duke@435: enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; duke@435: duke@435: switch (id) { duke@435: case vmIntrinsics::_isInstance: duke@435: nargs = 1+1; // the Class mirror, plus the object getting queried about duke@435: // nothing is an instance of a primitive type duke@435: prim_return_value = intcon(0); duke@435: break; duke@435: case vmIntrinsics::_getModifiers: duke@435: prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); duke@435: assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); duke@435: return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); duke@435: break; duke@435: case vmIntrinsics::_isInterface: duke@435: prim_return_value = intcon(0); duke@435: break; duke@435: case vmIntrinsics::_isArray: duke@435: prim_return_value = intcon(0); duke@435: expect_prim = true; // cf. ObjectStreamClass.getClassSignature duke@435: break; duke@435: case vmIntrinsics::_isPrimitive: duke@435: prim_return_value = intcon(1); duke@435: expect_prim = true; // obviously duke@435: break; duke@435: case vmIntrinsics::_getSuperclass: duke@435: prim_return_value = null(); duke@435: return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); duke@435: break; duke@435: case vmIntrinsics::_getComponentType: duke@435: prim_return_value = null(); duke@435: return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); duke@435: break; duke@435: case vmIntrinsics::_getClassAccessFlags: duke@435: prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); duke@435: return_type = TypeInt::INT; // not bool! 6297094 duke@435: break; duke@435: default: duke@435: ShouldNotReachHere(); duke@435: } duke@435: duke@435: Node* mirror = argument(0); duke@435: Node* obj = (nargs <= 1)? top(): argument(1); duke@435: duke@435: const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); duke@435: if (mirror_con == NULL) return false; // cannot happen? duke@435: duke@435: #ifndef PRODUCT duke@435: if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { duke@435: ciType* k = mirror_con->java_mirror_type(); duke@435: if (k) { duke@435: tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); duke@435: k->print_name(); duke@435: tty->cr(); duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). duke@435: RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); duke@435: record_for_igvn(region); duke@435: PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type); duke@435: duke@435: // The mirror will never be null of Reflection.getClassAccessFlags, however duke@435: // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE duke@435: // if it is. See bug 4774291. duke@435: duke@435: // For Reflection.getClassAccessFlags(), the null check occurs in duke@435: // the wrong place; see inline_unsafe_access(), above, for a similar duke@435: // situation. duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: mirror = do_null_check(mirror, T_OBJECT); duke@435: _sp -= nargs; duke@435: // If mirror or obj is dead, only null-path is taken. duke@435: if (stopped()) return true; duke@435: duke@435: if (expect_prim) never_see_null = false; // expect nulls (meaning prims) duke@435: duke@435: // Now load the mirror's klass metaobject, and null-check it. duke@435: // Side-effects region with the control path if the klass is null. duke@435: Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, duke@435: region, _prim_path); duke@435: // If kls is null, we have a primitive mirror. duke@435: phi->init_req(_prim_path, prim_return_value); duke@435: if (stopped()) { push_result(region, phi); return true; } duke@435: duke@435: Node* p; // handy temp duke@435: Node* null_ctl; duke@435: duke@435: // Now that we have the non-null klass, we can perform the real query. duke@435: // For constant classes, the query will constant-fold in LoadNode::Value. duke@435: Node* query_value = top(); duke@435: switch (id) { duke@435: case vmIntrinsics::_isInstance: duke@435: // nothing is an instance of a primitive type duke@435: query_value = gen_instanceof(obj, kls); duke@435: break; duke@435: duke@435: case vmIntrinsics::_getModifiers: duke@435: p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc)); duke@435: query_value = make_load(NULL, p, TypeInt::INT, T_INT); duke@435: break; duke@435: duke@435: case vmIntrinsics::_isInterface: duke@435: // (To verify this code sequence, check the asserts in JVM_IsInterface.) duke@435: if (generate_interface_guard(kls, region) != NULL) duke@435: // A guard was added. If the guard is taken, it was an interface. duke@435: phi->add_req(intcon(1)); duke@435: // If we fall through, it's a plain class. duke@435: query_value = intcon(0); duke@435: break; duke@435: duke@435: case vmIntrinsics::_isArray: duke@435: // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) duke@435: if (generate_array_guard(kls, region) != NULL) duke@435: // A guard was added. If the guard is taken, it was an array. duke@435: phi->add_req(intcon(1)); duke@435: // If we fall through, it's a plain class. duke@435: query_value = intcon(0); duke@435: break; duke@435: duke@435: case vmIntrinsics::_isPrimitive: duke@435: query_value = intcon(0); // "normal" path produces false duke@435: break; duke@435: duke@435: case vmIntrinsics::_getSuperclass: duke@435: // The rules here are somewhat unfortunate, but we can still do better duke@435: // with random logic than with a JNI call. duke@435: // Interfaces store null or Object as _super, but must report null. duke@435: // Arrays store an intermediate super as _super, but must report Object. duke@435: // Other types can report the actual _super. duke@435: // (To verify this code sequence, check the asserts in JVM_IsInterface.) duke@435: if (generate_interface_guard(kls, region) != NULL) duke@435: // A guard was added. If the guard is taken, it was an interface. duke@435: phi->add_req(null()); duke@435: if (generate_array_guard(kls, region) != NULL) duke@435: // A guard was added. If the guard is taken, it was an array. duke@435: phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); duke@435: // If we fall through, it's a plain class. Get its _super. duke@435: p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc)); kvn@599: kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) ); duke@435: null_ctl = top(); duke@435: kls = null_check_oop(kls, &null_ctl); duke@435: if (null_ctl != top()) { duke@435: // If the guard is taken, Object.superClass is null (both klass and mirror). duke@435: region->add_req(null_ctl); duke@435: phi ->add_req(null()); duke@435: } duke@435: if (!stopped()) { duke@435: query_value = load_mirror_from_klass(kls); duke@435: } duke@435: break; duke@435: duke@435: case vmIntrinsics::_getComponentType: duke@435: if (generate_array_guard(kls, region) != NULL) { duke@435: // Be sure to pin the oop load to the guard edge just created: duke@435: Node* is_array_ctrl = region->in(region->req()-1); duke@435: Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc)); duke@435: Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); duke@435: phi->add_req(cmo); duke@435: } duke@435: query_value = null(); // non-array case is null duke@435: break; duke@435: duke@435: case vmIntrinsics::_getClassAccessFlags: duke@435: p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); duke@435: query_value = make_load(NULL, p, TypeInt::INT, T_INT); duke@435: break; duke@435: duke@435: default: duke@435: ShouldNotReachHere(); duke@435: } duke@435: duke@435: // Fall-through is the normal case of a query to a real class. duke@435: phi->init_req(1, query_value); duke@435: region->init_req(1, control()); duke@435: duke@435: push_result(region, phi); duke@435: C->set_has_split_ifs(true); // Has chance for split-if optimization duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //--------------------------inline_native_subtype_check------------------------ duke@435: // This intrinsic takes the JNI calls out of the heart of duke@435: // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. duke@435: bool LibraryCallKit::inline_native_subtype_check() { duke@435: int nargs = 1+1; // the Class mirror, plus the other class getting examined duke@435: duke@435: // Pull both arguments off the stack. duke@435: Node* args[2]; // two java.lang.Class mirrors: superc, subc duke@435: args[0] = argument(0); duke@435: args[1] = argument(1); duke@435: Node* klasses[2]; // corresponding Klasses: superk, subk duke@435: klasses[0] = klasses[1] = top(); duke@435: duke@435: enum { duke@435: // A full decision tree on {superc is prim, subc is prim}: duke@435: _prim_0_path = 1, // {P,N} => false duke@435: // {P,P} & superc!=subc => false duke@435: _prim_same_path, // {P,P} & superc==subc => true duke@435: _prim_1_path, // {N,P} => false duke@435: _ref_subtype_path, // {N,N} & subtype check wins => true duke@435: _both_ref_path, // {N,N} & subtype check loses => false duke@435: PATH_LIMIT duke@435: }; duke@435: duke@435: RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); duke@435: Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); duke@435: record_for_igvn(region); duke@435: duke@435: const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads duke@435: const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; duke@435: int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); duke@435: duke@435: // First null-check both mirrors and load each mirror's klass metaobject. duke@435: int which_arg; duke@435: for (which_arg = 0; which_arg <= 1; which_arg++) { duke@435: Node* arg = args[which_arg]; duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: arg = do_null_check(arg, T_OBJECT); duke@435: _sp -= nargs; duke@435: if (stopped()) break; duke@435: args[which_arg] = _gvn.transform(arg); duke@435: duke@435: Node* p = basic_plus_adr(arg, class_klass_offset); kvn@599: Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); duke@435: klasses[which_arg] = _gvn.transform(kls); duke@435: } duke@435: duke@435: // Having loaded both klasses, test each for null. duke@435: bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); duke@435: for (which_arg = 0; which_arg <= 1; which_arg++) { duke@435: Node* kls = klasses[which_arg]; duke@435: Node* null_ctl = top(); duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: kls = null_check_oop(kls, &null_ctl, never_see_null); duke@435: _sp -= nargs; duke@435: int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); duke@435: region->init_req(prim_path, null_ctl); duke@435: if (stopped()) break; duke@435: klasses[which_arg] = kls; duke@435: } duke@435: duke@435: if (!stopped()) { duke@435: // now we have two reference types, in klasses[0..1] duke@435: Node* subk = klasses[1]; // the argument to isAssignableFrom duke@435: Node* superk = klasses[0]; // the receiver duke@435: region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); duke@435: // now we have a successful reference subtype check duke@435: region->set_req(_ref_subtype_path, control()); duke@435: } duke@435: duke@435: // If both operands are primitive (both klasses null), then duke@435: // we must return true when they are identical primitives. duke@435: // It is convenient to test this after the first null klass check. duke@435: set_control(region->in(_prim_0_path)); // go back to first null check duke@435: if (!stopped()) { duke@435: // Since superc is primitive, make a guard for the superc==subc case. duke@435: Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) ); duke@435: Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) ); duke@435: generate_guard(bol_eq, region, PROB_FAIR); duke@435: if (region->req() == PATH_LIMIT+1) { duke@435: // A guard was added. If the added guard is taken, superc==subc. duke@435: region->swap_edges(PATH_LIMIT, _prim_same_path); duke@435: region->del_req(PATH_LIMIT); duke@435: } duke@435: region->set_req(_prim_0_path, control()); // Not equal after all. duke@435: } duke@435: duke@435: // these are the only paths that produce 'true': duke@435: phi->set_req(_prim_same_path, intcon(1)); duke@435: phi->set_req(_ref_subtype_path, intcon(1)); duke@435: duke@435: // pull together the cases: duke@435: assert(region->req() == PATH_LIMIT, "sane region"); duke@435: for (uint i = 1; i < region->req(); i++) { duke@435: Node* ctl = region->in(i); duke@435: if (ctl == NULL || ctl == top()) { duke@435: region->set_req(i, top()); duke@435: phi ->set_req(i, top()); duke@435: } else if (phi->in(i) == NULL) { duke@435: phi->set_req(i, intcon(0)); // all other paths produce 'false' duke@435: } duke@435: } duke@435: duke@435: set_control(_gvn.transform(region)); duke@435: push(_gvn.transform(phi)); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //---------------------generate_array_guard_common------------------------ duke@435: Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, duke@435: bool obj_array, bool not_array) { duke@435: // If obj_array/non_array==false/false: duke@435: // Branch around if the given klass is in fact an array (either obj or prim). duke@435: // If obj_array/non_array==false/true: duke@435: // Branch around if the given klass is not an array klass of any kind. duke@435: // If obj_array/non_array==true/true: duke@435: // Branch around if the kls is not an oop array (kls is int[], String, etc.) duke@435: // If obj_array/non_array==true/false: duke@435: // Branch around if the kls is an oop array (Object[] or subtype) duke@435: // duke@435: // Like generate_guard, adds a new path onto the region. duke@435: jint layout_con = 0; duke@435: Node* layout_val = get_layout_helper(kls, layout_con); duke@435: if (layout_val == NULL) { duke@435: bool query = (obj_array duke@435: ? Klass::layout_helper_is_objArray(layout_con) duke@435: : Klass::layout_helper_is_javaArray(layout_con)); duke@435: if (query == not_array) { duke@435: return NULL; // never a branch duke@435: } else { // always a branch duke@435: Node* always_branch = control(); duke@435: if (region != NULL) duke@435: region->add_req(always_branch); duke@435: set_control(top()); duke@435: return always_branch; duke@435: } duke@435: } duke@435: // Now test the correct condition. duke@435: jint nval = (obj_array duke@435: ? ((jint)Klass::_lh_array_tag_type_value duke@435: << Klass::_lh_array_tag_shift) duke@435: : Klass::_lh_neutral_value); duke@435: Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) ); duke@435: BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array duke@435: // invert the test if we are looking for a non-array duke@435: if (not_array) btest = BoolTest(btest).negate(); duke@435: Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) ); duke@435: return generate_fair_guard(bol, region); duke@435: } duke@435: duke@435: duke@435: //-----------------------inline_native_newArray-------------------------- duke@435: bool LibraryCallKit::inline_native_newArray() { duke@435: int nargs = 2; duke@435: Node* mirror = argument(0); duke@435: Node* count_val = argument(1); duke@435: duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: mirror = do_null_check(mirror, T_OBJECT); duke@435: _sp -= nargs; kvn@598: // If mirror or obj is dead, only null-path is taken. kvn@598: if (stopped()) return true; duke@435: duke@435: enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; duke@435: RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); duke@435: PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, duke@435: TypeInstPtr::NOTNULL); duke@435: PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); duke@435: PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, duke@435: TypePtr::BOTTOM); duke@435: duke@435: bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); duke@435: Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, duke@435: nargs, duke@435: result_reg, _slow_path); duke@435: Node* normal_ctl = control(); duke@435: Node* no_array_ctl = result_reg->in(_slow_path); duke@435: duke@435: // Generate code for the slow case. We make a call to newArray(). duke@435: set_control(no_array_ctl); duke@435: if (!stopped()) { duke@435: // Either the input type is void.class, or else the duke@435: // array klass has not yet been cached. Either the duke@435: // ensuing call will throw an exception, or else it duke@435: // will cache the array klass for next time. duke@435: PreserveJVMState pjvms(this); duke@435: CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); duke@435: Node* slow_result = set_results_for_java_call(slow_call); duke@435: // this->control() comes from set_results_for_java_call duke@435: result_reg->set_req(_slow_path, control()); duke@435: result_val->set_req(_slow_path, slow_result); duke@435: result_io ->set_req(_slow_path, i_o()); duke@435: result_mem->set_req(_slow_path, reset_memory()); duke@435: } duke@435: duke@435: set_control(normal_ctl); duke@435: if (!stopped()) { duke@435: // Normal case: The array type has been cached in the java.lang.Class. duke@435: // The following call works fine even if the array type is polymorphic. duke@435: // It could be a dynamic mix of int[], boolean[], Object[], etc. duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: Node* obj = new_array(klass_node, count_val); duke@435: _sp -= nargs; duke@435: result_reg->init_req(_normal_path, control()); duke@435: result_val->init_req(_normal_path, obj); duke@435: result_io ->init_req(_normal_path, i_o()); duke@435: result_mem->init_req(_normal_path, reset_memory()); duke@435: } duke@435: duke@435: // Return the combined state. duke@435: set_i_o( _gvn.transform(result_io) ); duke@435: set_all_memory( _gvn.transform(result_mem) ); duke@435: push_result(result_reg, result_val); duke@435: C->set_has_split_ifs(true); // Has chance for split-if optimization duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //----------------------inline_native_getLength-------------------------- duke@435: bool LibraryCallKit::inline_native_getLength() { duke@435: if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; duke@435: duke@435: int nargs = 1; duke@435: Node* array = argument(0); duke@435: duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: array = do_null_check(array, T_OBJECT); duke@435: _sp -= nargs; duke@435: duke@435: // If array is dead, only null-path is taken. duke@435: if (stopped()) return true; duke@435: duke@435: // Deoptimize if it is a non-array. duke@435: Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); duke@435: duke@435: if (non_array != NULL) { duke@435: PreserveJVMState pjvms(this); duke@435: set_control(non_array); duke@435: _sp += nargs; // push the arguments back on the stack duke@435: uncommon_trap(Deoptimization::Reason_intrinsic, duke@435: Deoptimization::Action_maybe_recompile); duke@435: } duke@435: duke@435: // If control is dead, only non-array-path is taken. duke@435: if (stopped()) return true; duke@435: duke@435: // The works fine even if the array type is polymorphic. duke@435: // It could be a dynamic mix of int[], boolean[], Object[], etc. duke@435: push( load_array_length(array) ); duke@435: duke@435: C->set_has_split_ifs(true); // Has chance for split-if optimization duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //------------------------inline_array_copyOf---------------------------- duke@435: bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { duke@435: if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; duke@435: duke@435: // Restore the stack and pop off the arguments. duke@435: int nargs = 3 + (is_copyOfRange? 1: 0); duke@435: Node* original = argument(0); duke@435: Node* start = is_copyOfRange? argument(1): intcon(0); duke@435: Node* end = is_copyOfRange? argument(2): argument(1); duke@435: Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); duke@435: duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); duke@435: original = do_null_check(original, T_OBJECT); duke@435: _sp -= nargs; duke@435: duke@435: // Check if a null path was taken unconditionally. duke@435: if (stopped()) return true; duke@435: duke@435: Node* orig_length = load_array_length(original); duke@435: duke@435: Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs, duke@435: NULL, 0); duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: klass_node = do_null_check(klass_node, T_OBJECT); duke@435: _sp -= nargs; duke@435: duke@435: RegionNode* bailout = new (C, 1) RegionNode(1); duke@435: record_for_igvn(bailout); duke@435: duke@435: // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. duke@435: // Bail out if that is so. duke@435: Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); duke@435: if (not_objArray != NULL) { duke@435: // Improve the klass node's type from the new optimistic assumption: duke@435: ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); duke@435: const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); duke@435: Node* cast = new (C, 2) CastPPNode(klass_node, akls); duke@435: cast->init_req(0, control()); duke@435: klass_node = _gvn.transform(cast); duke@435: } duke@435: duke@435: // Bail out if either start or end is negative. duke@435: generate_negative_guard(start, bailout, &start); duke@435: generate_negative_guard(end, bailout, &end); duke@435: duke@435: Node* length = end; duke@435: if (_gvn.type(start) != TypeInt::ZERO) { duke@435: length = _gvn.transform( new (C, 3) SubINode(end, start) ); duke@435: } duke@435: duke@435: // Bail out if length is negative. duke@435: // ...Not needed, since the new_array will throw the right exception. duke@435: //generate_negative_guard(length, bailout, &length); duke@435: duke@435: if (bailout->req() > 1) { duke@435: PreserveJVMState pjvms(this); duke@435: set_control( _gvn.transform(bailout) ); duke@435: _sp += nargs; // push the arguments back on the stack duke@435: uncommon_trap(Deoptimization::Reason_intrinsic, duke@435: Deoptimization::Action_maybe_recompile); duke@435: } duke@435: duke@435: if (!stopped()) { duke@435: // How many elements will we copy from the original? duke@435: // The answer is MinI(orig_length - start, length). duke@435: Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) ); duke@435: Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); duke@435: duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: Node* newcopy = new_array(klass_node, length); duke@435: _sp -= nargs; duke@435: duke@435: // Generate a direct call to the right arraycopy function(s). duke@435: // We know the copy is disjoint but we might not know if the duke@435: // oop stores need checking. duke@435: // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). duke@435: // This will fail a store-check if x contains any non-nulls. duke@435: bool disjoint_bases = true; duke@435: bool length_never_negative = true; duke@435: generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, duke@435: original, start, newcopy, intcon(0), moved, duke@435: nargs, disjoint_bases, length_never_negative); duke@435: duke@435: push(newcopy); duke@435: } duke@435: duke@435: C->set_has_split_ifs(true); // Has chance for split-if optimization duke@435: duke@435: return true; duke@435: } duke@435: duke@435: duke@435: //----------------------generate_virtual_guard--------------------------- duke@435: // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. duke@435: Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, duke@435: RegionNode* slow_region) { duke@435: ciMethod* method = callee(); duke@435: int vtable_index = method->vtable_index(); duke@435: // Get the methodOop out of the appropriate vtable entry. duke@435: int entry_offset = (instanceKlass::vtable_start_offset() + duke@435: vtable_index*vtableEntry::size()) * wordSize + duke@435: vtableEntry::method_offset_in_bytes(); duke@435: Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); duke@435: Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT); duke@435: duke@435: // Compare the target method with the expected method (e.g., Object.hashCode). duke@435: const TypeInstPtr* native_call_addr = TypeInstPtr::make(method); duke@435: duke@435: Node* native_call = makecon(native_call_addr); duke@435: Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) ); duke@435: Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) ); duke@435: duke@435: return generate_slow_guard(test_native, slow_region); duke@435: } duke@435: duke@435: //-----------------------generate_method_call---------------------------- duke@435: // Use generate_method_call to make a slow-call to the real duke@435: // method if the fast path fails. An alternative would be to duke@435: // use a stub like OptoRuntime::slow_arraycopy_Java. duke@435: // This only works for expanding the current library call, duke@435: // not another intrinsic. (E.g., don't use this for making an duke@435: // arraycopy call inside of the copyOf intrinsic.) duke@435: CallJavaNode* duke@435: LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { duke@435: // When compiling the intrinsic method itself, do not use this technique. duke@435: guarantee(callee() != C->method(), "cannot make slow-call to self"); duke@435: duke@435: ciMethod* method = callee(); duke@435: // ensure the JVMS we have will be correct for this call duke@435: guarantee(method_id == method->intrinsic_id(), "must match"); duke@435: duke@435: const TypeFunc* tf = TypeFunc::make(method); duke@435: int tfdc = tf->domain()->cnt(); duke@435: CallJavaNode* slow_call; duke@435: if (is_static) { duke@435: assert(!is_virtual, ""); duke@435: slow_call = new(C, tfdc) CallStaticJavaNode(tf, duke@435: SharedRuntime::get_resolve_static_call_stub(), duke@435: method, bci()); duke@435: } else if (is_virtual) { duke@435: null_check_receiver(method); duke@435: int vtable_index = methodOopDesc::invalid_vtable_index; duke@435: if (UseInlineCaches) { duke@435: // Suppress the vtable call duke@435: } else { duke@435: // hashCode and clone are not a miranda methods, duke@435: // so the vtable index is fixed. duke@435: // No need to use the linkResolver to get it. duke@435: vtable_index = method->vtable_index(); duke@435: } duke@435: slow_call = new(C, tfdc) CallDynamicJavaNode(tf, duke@435: SharedRuntime::get_resolve_virtual_call_stub(), duke@435: method, vtable_index, bci()); duke@435: } else { // neither virtual nor static: opt_virtual duke@435: null_check_receiver(method); duke@435: slow_call = new(C, tfdc) CallStaticJavaNode(tf, duke@435: SharedRuntime::get_resolve_opt_virtual_call_stub(), duke@435: method, bci()); duke@435: slow_call->set_optimized_virtual(true); duke@435: } duke@435: set_arguments_for_java_call(slow_call); duke@435: set_edges_for_java_call(slow_call); duke@435: return slow_call; duke@435: } duke@435: duke@435: duke@435: //------------------------------inline_native_hashcode-------------------- duke@435: // Build special case code for calls to hashCode on an object. duke@435: bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { duke@435: assert(is_static == callee()->is_static(), "correct intrinsic selection"); duke@435: assert(!(is_virtual && is_static), "either virtual, special, or static"); duke@435: duke@435: enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; duke@435: duke@435: RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); duke@435: PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, duke@435: TypeInt::INT); duke@435: PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); duke@435: PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, duke@435: TypePtr::BOTTOM); duke@435: Node* obj = NULL; duke@435: if (!is_static) { duke@435: // Check for hashing null object duke@435: obj = null_check_receiver(callee()); duke@435: if (stopped()) return true; // unconditionally null duke@435: result_reg->init_req(_null_path, top()); duke@435: result_val->init_req(_null_path, top()); duke@435: } else { duke@435: // Do a null check, and return zero if null. duke@435: // System.identityHashCode(null) == 0 duke@435: obj = argument(0); duke@435: Node* null_ctl = top(); duke@435: obj = null_check_oop(obj, &null_ctl); duke@435: result_reg->init_req(_null_path, null_ctl); duke@435: result_val->init_req(_null_path, _gvn.intcon(0)); duke@435: } duke@435: duke@435: // Unconditionally null? Then return right away. duke@435: if (stopped()) { duke@435: set_control( result_reg->in(_null_path) ); duke@435: if (!stopped()) duke@435: push( result_val ->in(_null_path) ); duke@435: return true; duke@435: } duke@435: duke@435: // After null check, get the object's klass. duke@435: Node* obj_klass = load_object_klass(obj); duke@435: duke@435: // This call may be virtual (invokevirtual) or bound (invokespecial). duke@435: // For each case we generate slightly different code. duke@435: duke@435: // We only go to the fast case code if we pass a number of guards. The duke@435: // paths which do not pass are accumulated in the slow_region. duke@435: RegionNode* slow_region = new (C, 1) RegionNode(1); duke@435: record_for_igvn(slow_region); duke@435: duke@435: // If this is a virtual call, we generate a funny guard. We pull out duke@435: // the vtable entry corresponding to hashCode() from the target object. duke@435: // If the target method which we are calling happens to be the native duke@435: // Object hashCode() method, we pass the guard. We do not need this duke@435: // guard for non-virtual calls -- the caller is known to be the native duke@435: // Object hashCode(). duke@435: if (is_virtual) { duke@435: generate_virtual_guard(obj_klass, slow_region); duke@435: } duke@435: duke@435: // Get the header out of the object, use LoadMarkNode when available duke@435: Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); duke@435: Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS); duke@435: header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) ); duke@435: duke@435: // Test the header to see if it is unlocked. duke@435: Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); duke@435: Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) ); duke@435: Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); duke@435: Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val)); duke@435: Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) ); duke@435: duke@435: generate_slow_guard(test_unlocked, slow_region); duke@435: duke@435: // Get the hash value and check to see that it has been properly assigned. duke@435: // We depend on hash_mask being at most 32 bits and avoid the use of duke@435: // hash_mask_in_place because it could be larger than 32 bits in a 64-bit duke@435: // vm: see markOop.hpp. duke@435: Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); duke@435: Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); duke@435: Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) ); duke@435: // This hack lets the hash bits live anywhere in the mark object now, as long twisti@1040: // as the shift drops the relevant bits into the low 32 bits. Note that duke@435: // Java spec says that HashCode is an int so there's no point in capturing duke@435: // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). duke@435: hshifted_header = ConvX2I(hshifted_header); duke@435: Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) ); duke@435: duke@435: Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); duke@435: Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val)); duke@435: Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) ); duke@435: duke@435: generate_slow_guard(test_assigned, slow_region); duke@435: duke@435: Node* init_mem = reset_memory(); duke@435: // fill in the rest of the null path: duke@435: result_io ->init_req(_null_path, i_o()); duke@435: result_mem->init_req(_null_path, init_mem); duke@435: duke@435: result_val->init_req(_fast_path, hash_val); duke@435: result_reg->init_req(_fast_path, control()); duke@435: result_io ->init_req(_fast_path, i_o()); duke@435: result_mem->init_req(_fast_path, init_mem); duke@435: duke@435: // Generate code for the slow case. We make a call to hashCode(). duke@435: set_control(_gvn.transform(slow_region)); duke@435: if (!stopped()) { duke@435: // No need for PreserveJVMState, because we're using up the present state. duke@435: set_all_memory(init_mem); duke@435: vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; duke@435: if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; duke@435: CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); duke@435: Node* slow_result = set_results_for_java_call(slow_call); duke@435: // this->control() comes from set_results_for_java_call duke@435: result_reg->init_req(_slow_path, control()); duke@435: result_val->init_req(_slow_path, slow_result); duke@435: result_io ->set_req(_slow_path, i_o()); duke@435: result_mem ->set_req(_slow_path, reset_memory()); duke@435: } duke@435: duke@435: // Return the combined state. duke@435: set_i_o( _gvn.transform(result_io) ); duke@435: set_all_memory( _gvn.transform(result_mem) ); duke@435: push_result(result_reg, result_val); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //---------------------------inline_native_getClass---------------------------- twisti@1040: // Build special case code for calls to getClass on an object. duke@435: bool LibraryCallKit::inline_native_getClass() { duke@435: Node* obj = null_check_receiver(callee()); duke@435: if (stopped()) return true; duke@435: push( load_mirror_from_klass(load_object_klass(obj)) ); duke@435: return true; duke@435: } duke@435: duke@435: //-----------------inline_native_Reflection_getCallerClass--------------------- duke@435: // In the presence of deep enough inlining, getCallerClass() becomes a no-op. duke@435: // duke@435: // NOTE that this code must perform the same logic as duke@435: // vframeStream::security_get_caller_frame in that it must skip duke@435: // Method.invoke() and auxiliary frames. duke@435: duke@435: duke@435: duke@435: duke@435: bool LibraryCallKit::inline_native_Reflection_getCallerClass() { duke@435: ciMethod* method = callee(); duke@435: duke@435: #ifndef PRODUCT duke@435: if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { duke@435: tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); duke@435: } duke@435: #endif duke@435: duke@435: debug_only(int saved_sp = _sp); duke@435: duke@435: // Argument words: (int depth) duke@435: int nargs = 1; duke@435: duke@435: _sp += nargs; duke@435: Node* caller_depth_node = pop(); duke@435: duke@435: assert(saved_sp == _sp, "must have correct argument count"); duke@435: duke@435: // The depth value must be a constant in order for the runtime call duke@435: // to be eliminated. duke@435: const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); duke@435: if (caller_depth_type == NULL || !caller_depth_type->is_con()) { duke@435: #ifndef PRODUCT duke@435: if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { duke@435: tty->print_cr(" Bailing out because caller depth was not a constant"); duke@435: } duke@435: #endif duke@435: return false; duke@435: } duke@435: // Note that the JVM state at this point does not include the duke@435: // getCallerClass() frame which we are trying to inline. The duke@435: // semantics of getCallerClass(), however, are that the "first" duke@435: // frame is the getCallerClass() frame, so we subtract one from the duke@435: // requested depth before continuing. We don't inline requests of duke@435: // getCallerClass(0). duke@435: int caller_depth = caller_depth_type->get_con() - 1; duke@435: if (caller_depth < 0) { duke@435: #ifndef PRODUCT duke@435: if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { duke@435: tty->print_cr(" Bailing out because caller depth was %d", caller_depth); duke@435: } duke@435: #endif duke@435: return false; duke@435: } duke@435: duke@435: if (!jvms()->has_method()) { duke@435: #ifndef PRODUCT duke@435: if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { duke@435: tty->print_cr(" Bailing out because intrinsic was inlined at top level"); duke@435: } duke@435: #endif duke@435: return false; duke@435: } duke@435: int _depth = jvms()->depth(); // cache call chain depth duke@435: duke@435: // Walk back up the JVM state to find the caller at the required duke@435: // depth. NOTE that this code must perform the same logic as duke@435: // vframeStream::security_get_caller_frame in that it must skip duke@435: // Method.invoke() and auxiliary frames. Note also that depth is duke@435: // 1-based (1 is the bottom of the inlining). duke@435: int inlining_depth = _depth; duke@435: JVMState* caller_jvms = NULL; duke@435: duke@435: if (inlining_depth > 0) { duke@435: caller_jvms = jvms(); duke@435: assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); duke@435: do { duke@435: // The following if-tests should be performed in this order duke@435: if (is_method_invoke_or_aux_frame(caller_jvms)) { duke@435: // Skip a Method.invoke() or auxiliary frame duke@435: } else if (caller_depth > 0) { duke@435: // Skip real frame duke@435: --caller_depth; duke@435: } else { duke@435: // We're done: reached desired caller after skipping. duke@435: break; duke@435: } duke@435: caller_jvms = caller_jvms->caller(); duke@435: --inlining_depth; duke@435: } while (inlining_depth > 0); duke@435: } duke@435: duke@435: if (inlining_depth == 0) { duke@435: #ifndef PRODUCT duke@435: if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { duke@435: tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); duke@435: tty->print_cr(" JVM state at this point:"); duke@435: for (int i = _depth; i >= 1; i--) { duke@435: tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); duke@435: } duke@435: } duke@435: #endif duke@435: return false; // Reached end of inlining duke@435: } duke@435: duke@435: // Acquire method holder as java.lang.Class duke@435: ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); duke@435: ciInstance* caller_mirror = caller_klass->java_mirror(); duke@435: // Push this as a constant duke@435: push(makecon(TypeInstPtr::make(caller_mirror))); duke@435: #ifndef PRODUCT duke@435: if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { duke@435: tty->print_cr(" Succeeded: caller = %s.%s, caller depth = %d, depth = %d", caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), caller_depth_type->get_con(), _depth); duke@435: tty->print_cr(" JVM state at this point:"); duke@435: for (int i = _depth; i >= 1; i--) { duke@435: tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); duke@435: } duke@435: } duke@435: #endif duke@435: return true; duke@435: } duke@435: duke@435: // Helper routine for above duke@435: bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { duke@435: // Is this the Method.invoke method itself? duke@435: if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke) duke@435: return true; duke@435: duke@435: // Is this a helper, defined somewhere underneath MethodAccessorImpl. duke@435: ciKlass* k = jvms->method()->holder(); duke@435: if (k->is_instance_klass()) { duke@435: ciInstanceKlass* ik = k->as_instance_klass(); duke@435: for (; ik != NULL; ik = ik->super()) { duke@435: if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && duke@435: ik == env()->find_system_klass(ik->name())) { duke@435: return true; duke@435: } duke@435: } duke@435: } duke@435: duke@435: return false; duke@435: } duke@435: duke@435: static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by duke@435: // inline_native_AtomicLong_attemptUpdate() but it has no way of duke@435: // computing it since there is no lookup field by name function in the duke@435: // CI interface. This is computed and set by inline_native_AtomicLong_get(). duke@435: // Using a static variable here is safe even if we have multiple compilation duke@435: // threads because the offset is constant. At worst the same offset will be duke@435: // computed and stored multiple duke@435: duke@435: bool LibraryCallKit::inline_native_AtomicLong_get() { duke@435: // Restore the stack and pop off the argument duke@435: _sp+=1; duke@435: Node *obj = pop(); duke@435: duke@435: // get the offset of the "value" field. Since the CI interfaces duke@435: // does not provide a way to look up a field by name, we scan the bytecodes duke@435: // to get the field index. We expect the first 2 instructions of the method duke@435: // to be: duke@435: // 0 aload_0 duke@435: // 1 getfield "value" duke@435: ciMethod* method = callee(); duke@435: if (value_field_offset == -1) duke@435: { duke@435: ciField* value_field; duke@435: ciBytecodeStream iter(method); duke@435: Bytecodes::Code bc = iter.next(); duke@435: duke@435: if ((bc != Bytecodes::_aload_0) && duke@435: ((bc != Bytecodes::_aload) || (iter.get_index() != 0))) duke@435: return false; duke@435: bc = iter.next(); duke@435: if (bc != Bytecodes::_getfield) duke@435: return false; duke@435: bool ignore; duke@435: value_field = iter.get_field(ignore); duke@435: value_field_offset = value_field->offset_in_bytes(); duke@435: } duke@435: duke@435: // Null check without removing any arguments. duke@435: _sp++; duke@435: obj = do_null_check(obj, T_OBJECT); duke@435: _sp--; duke@435: // Check for locking null object duke@435: if (stopped()) return true; duke@435: duke@435: Node *adr = basic_plus_adr(obj, obj, value_field_offset); duke@435: const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); duke@435: int alias_idx = C->get_alias_index(adr_type); duke@435: duke@435: Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr)); duke@435: duke@435: push_pair(result); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() { duke@435: // Restore the stack and pop off the arguments duke@435: _sp+=5; duke@435: Node *newVal = pop_pair(); duke@435: Node *oldVal = pop_pair(); duke@435: Node *obj = pop(); duke@435: duke@435: // we need the offset of the "value" field which was computed when duke@435: // inlining the get() method. Give up if we don't have it. duke@435: if (value_field_offset == -1) duke@435: return false; duke@435: duke@435: // Null check without removing any arguments. duke@435: _sp+=5; duke@435: obj = do_null_check(obj, T_OBJECT); duke@435: _sp-=5; duke@435: // Check for locking null object duke@435: if (stopped()) return true; duke@435: duke@435: Node *adr = basic_plus_adr(obj, obj, value_field_offset); duke@435: const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); duke@435: int alias_idx = C->get_alias_index(adr_type); duke@435: kvn@855: Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal)); kvn@855: Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); duke@435: set_memory(store_proj, alias_idx); kvn@855: Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) ); kvn@855: kvn@855: Node *result; kvn@855: // CMove node is not used to be able fold a possible check code kvn@855: // after attemptUpdate() call. This code could be transformed kvn@855: // into CMove node by loop optimizations. kvn@855: { kvn@855: RegionNode *r = new (C, 3) RegionNode(3); kvn@855: result = new (C, 3) PhiNode(r, TypeInt::BOOL); kvn@855: kvn@855: Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); kvn@855: Node *iftrue = opt_iff(r, iff); kvn@855: r->init_req(1, iftrue); kvn@855: result->init_req(1, intcon(1)); kvn@855: result->init_req(2, intcon(0)); kvn@855: kvn@855: set_control(_gvn.transform(r)); kvn@855: record_for_igvn(r); kvn@855: kvn@855: C->set_has_split_ifs(true); // Has chance for split-if optimization kvn@855: } kvn@855: kvn@855: push(_gvn.transform(result)); duke@435: return true; duke@435: } duke@435: duke@435: bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { duke@435: // restore the arguments duke@435: _sp += arg_size(); duke@435: duke@435: switch (id) { duke@435: case vmIntrinsics::_floatToRawIntBits: duke@435: push(_gvn.transform( new (C, 2) MoveF2INode(pop()))); duke@435: break; duke@435: duke@435: case vmIntrinsics::_intBitsToFloat: duke@435: push(_gvn.transform( new (C, 2) MoveI2FNode(pop()))); duke@435: break; duke@435: duke@435: case vmIntrinsics::_doubleToRawLongBits: duke@435: push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair()))); duke@435: break; duke@435: duke@435: case vmIntrinsics::_longBitsToDouble: duke@435: push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair()))); duke@435: break; duke@435: duke@435: case vmIntrinsics::_doubleToLongBits: { duke@435: Node* value = pop_pair(); duke@435: duke@435: // two paths (plus control) merge in a wood duke@435: RegionNode *r = new (C, 3) RegionNode(3); duke@435: Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); duke@435: duke@435: Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value)); duke@435: // Build the boolean node duke@435: Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); duke@435: duke@435: // Branch either way. duke@435: // NaN case is less traveled, which makes all the difference. duke@435: IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); duke@435: Node *opt_isnan = _gvn.transform(ifisnan); duke@435: assert( opt_isnan->is_If(), "Expect an IfNode"); duke@435: IfNode *opt_ifisnan = (IfNode*)opt_isnan; duke@435: Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); duke@435: duke@435: set_control(iftrue); duke@435: duke@435: static const jlong nan_bits = CONST64(0x7ff8000000000000); duke@435: Node *slow_result = longcon(nan_bits); // return NaN duke@435: phi->init_req(1, _gvn.transform( slow_result )); duke@435: r->init_req(1, iftrue); duke@435: duke@435: // Else fall through duke@435: Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); duke@435: set_control(iffalse); duke@435: duke@435: phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value))); duke@435: r->init_req(2, iffalse); duke@435: duke@435: // Post merge duke@435: set_control(_gvn.transform(r)); duke@435: record_for_igvn(r); duke@435: duke@435: Node* result = _gvn.transform(phi); duke@435: assert(result->bottom_type()->isa_long(), "must be"); duke@435: push_pair(result); duke@435: duke@435: C->set_has_split_ifs(true); // Has chance for split-if optimization duke@435: duke@435: break; duke@435: } duke@435: duke@435: case vmIntrinsics::_floatToIntBits: { duke@435: Node* value = pop(); duke@435: duke@435: // two paths (plus control) merge in a wood duke@435: RegionNode *r = new (C, 3) RegionNode(3); duke@435: Node *phi = new (C, 3) PhiNode(r, TypeInt::INT); duke@435: duke@435: Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value)); duke@435: // Build the boolean node duke@435: Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); duke@435: duke@435: // Branch either way. duke@435: // NaN case is less traveled, which makes all the difference. duke@435: IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); duke@435: Node *opt_isnan = _gvn.transform(ifisnan); duke@435: assert( opt_isnan->is_If(), "Expect an IfNode"); duke@435: IfNode *opt_ifisnan = (IfNode*)opt_isnan; duke@435: Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); duke@435: duke@435: set_control(iftrue); duke@435: duke@435: static const jint nan_bits = 0x7fc00000; duke@435: Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN duke@435: phi->init_req(1, _gvn.transform( slow_result )); duke@435: r->init_req(1, iftrue); duke@435: duke@435: // Else fall through duke@435: Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); duke@435: set_control(iffalse); duke@435: duke@435: phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value))); duke@435: r->init_req(2, iffalse); duke@435: duke@435: // Post merge duke@435: set_control(_gvn.transform(r)); duke@435: record_for_igvn(r); duke@435: duke@435: Node* result = _gvn.transform(phi); duke@435: assert(result->bottom_type()->isa_int(), "must be"); duke@435: push(result); duke@435: duke@435: C->set_has_split_ifs(true); // Has chance for split-if optimization duke@435: duke@435: break; duke@435: } duke@435: duke@435: default: duke@435: ShouldNotReachHere(); duke@435: } duke@435: duke@435: return true; duke@435: } duke@435: duke@435: #ifdef _LP64 duke@435: #define XTOP ,top() /*additional argument*/ duke@435: #else //_LP64 duke@435: #define XTOP /*no additional argument*/ duke@435: #endif //_LP64 duke@435: duke@435: //----------------------inline_unsafe_copyMemory------------------------- duke@435: bool LibraryCallKit::inline_unsafe_copyMemory() { duke@435: if (callee()->is_static()) return false; // caller must have the capability! duke@435: int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) duke@435: assert(signature()->size() == nargs-1, "copy has 5 arguments"); duke@435: null_check_receiver(callee()); // check then ignore argument(0) duke@435: if (stopped()) return true; duke@435: duke@435: C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". duke@435: duke@435: Node* src_ptr = argument(1); duke@435: Node* src_off = ConvL2X(argument(2)); duke@435: assert(argument(3)->is_top(), "2nd half of long"); duke@435: Node* dst_ptr = argument(4); duke@435: Node* dst_off = ConvL2X(argument(5)); duke@435: assert(argument(6)->is_top(), "2nd half of long"); duke@435: Node* size = ConvL2X(argument(7)); duke@435: assert(argument(8)->is_top(), "2nd half of long"); duke@435: duke@435: assert(Unsafe_field_offset_to_byte_offset(11) == 11, duke@435: "fieldOffset must be byte-scaled"); duke@435: duke@435: Node* src = make_unsafe_address(src_ptr, src_off); duke@435: Node* dst = make_unsafe_address(dst_ptr, dst_off); duke@435: duke@435: // Conservatively insert a memory barrier on all memory slices. duke@435: // Do not let writes of the copy source or destination float below the copy. duke@435: insert_mem_bar(Op_MemBarCPUOrder); duke@435: duke@435: // Call it. Note that the length argument is not scaled. duke@435: make_runtime_call(RC_LEAF|RC_NO_FP, duke@435: OptoRuntime::fast_arraycopy_Type(), duke@435: StubRoutines::unsafe_arraycopy(), duke@435: "unsafe_arraycopy", duke@435: TypeRawPtr::BOTTOM, duke@435: src, dst, size XTOP); duke@435: duke@435: // Do not let reads of the copy destination float above the copy. duke@435: insert_mem_bar(Op_MemBarCPUOrder); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: duke@435: //------------------------inline_native_clone---------------------------- duke@435: // Here are the simple edge cases: duke@435: // null receiver => normal trap duke@435: // virtual and clone was overridden => slow path to out-of-line clone duke@435: // not cloneable or finalizer => slow path to out-of-line Object.clone duke@435: // duke@435: // The general case has two steps, allocation and copying. duke@435: // Allocation has two cases, and uses GraphKit::new_instance or new_array. duke@435: // duke@435: // Copying also has two cases, oop arrays and everything else. duke@435: // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). duke@435: // Everything else uses the tight inline loop supplied by CopyArrayNode. duke@435: // duke@435: // These steps fold up nicely if and when the cloned object's klass duke@435: // can be sharply typed as an object array, a type array, or an instance. duke@435: // duke@435: bool LibraryCallKit::inline_native_clone(bool is_virtual) { duke@435: int nargs = 1; duke@435: Node* obj = null_check_receiver(callee()); duke@435: if (stopped()) return true; duke@435: Node* obj_klass = load_object_klass(obj); duke@435: const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); duke@435: const TypeOopPtr* toop = ((tklass != NULL) duke@435: ? tklass->as_instance_type() duke@435: : TypeInstPtr::NOTNULL); duke@435: duke@435: // Conservatively insert a memory barrier on all memory slices. duke@435: // Do not let writes into the original float below the clone. duke@435: insert_mem_bar(Op_MemBarCPUOrder); duke@435: duke@435: // paths into result_reg: duke@435: enum { duke@435: _slow_path = 1, // out-of-line call to clone method (virtual or not) duke@435: _objArray_path, // plain allocation, plus arrayof_oop_arraycopy duke@435: _fast_path, // plain allocation, plus a CopyArray operation duke@435: PATH_LIMIT duke@435: }; duke@435: RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); duke@435: PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, duke@435: TypeInstPtr::NOTNULL); duke@435: PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); duke@435: PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, duke@435: TypePtr::BOTTOM); duke@435: record_for_igvn(result_reg); duke@435: duke@435: const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; duke@435: int raw_adr_idx = Compile::AliasIdxRaw; duke@435: const bool raw_mem_only = true; duke@435: duke@435: // paths into alloc_reg (on the fast path, just before the CopyArray): duke@435: enum { _typeArray_alloc = 1, _instance_alloc, ALLOC_LIMIT }; duke@435: RegionNode* alloc_reg = new(C, ALLOC_LIMIT) RegionNode(ALLOC_LIMIT); duke@435: PhiNode* alloc_val = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, raw_adr_type); duke@435: PhiNode* alloc_siz = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, TypeX_X); duke@435: PhiNode* alloc_i_o = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::ABIO); duke@435: PhiNode* alloc_mem = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::MEMORY, duke@435: raw_adr_type); duke@435: record_for_igvn(alloc_reg); duke@435: duke@435: bool card_mark = false; // (see below) duke@435: duke@435: Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); duke@435: if (array_ctl != NULL) { duke@435: // It's an array. duke@435: PreserveJVMState pjvms(this); duke@435: set_control(array_ctl); duke@435: Node* obj_length = load_array_length(obj); duke@435: Node* obj_size = NULL; duke@435: _sp += nargs; // set original stack for use by uncommon_trap duke@435: Node* alloc_obj = new_array(obj_klass, obj_length, duke@435: raw_mem_only, &obj_size); duke@435: _sp -= nargs; duke@435: assert(obj_size != NULL, ""); duke@435: Node* raw_obj = alloc_obj->in(1); duke@435: assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); duke@435: if (ReduceBulkZeroing) { duke@435: AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); duke@435: if (alloc != NULL) { duke@435: // We will be completely responsible for initializing this object. duke@435: alloc->maybe_set_complete(&_gvn); duke@435: } duke@435: } duke@435: duke@435: if (!use_ReduceInitialCardMarks()) { duke@435: // If it is an oop array, it requires very special treatment, duke@435: // because card marking is required on each card of the array. duke@435: Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); duke@435: if (is_obja != NULL) { duke@435: PreserveJVMState pjvms2(this); duke@435: set_control(is_obja); duke@435: // Generate a direct call to the right arraycopy function(s). duke@435: bool disjoint_bases = true; duke@435: bool length_never_negative = true; duke@435: generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, duke@435: obj, intcon(0), alloc_obj, intcon(0), duke@435: obj_length, nargs, duke@435: disjoint_bases, length_never_negative); duke@435: result_reg->init_req(_objArray_path, control()); duke@435: result_val->init_req(_objArray_path, alloc_obj); duke@435: result_i_o ->set_req(_objArray_path, i_o()); duke@435: result_mem ->set_req(_objArray_path, reset_memory()); duke@435: } duke@435: } duke@435: // We can dispense with card marks if we know the allocation duke@435: // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks duke@435: // causes the non-eden paths to simulate a fresh allocation, duke@435: // insofar that no further card marks are required to initialize duke@435: // the object. duke@435: duke@435: // Otherwise, there are no card marks to worry about. duke@435: alloc_val->init_req(_typeArray_alloc, raw_obj); duke@435: alloc_siz->init_req(_typeArray_alloc, obj_size); duke@435: alloc_reg->init_req(_typeArray_alloc, control()); duke@435: alloc_i_o->init_req(_typeArray_alloc, i_o()); duke@435: alloc_mem->init_req(_typeArray_alloc, memory(raw_adr_type)); duke@435: } duke@435: duke@435: // We only go to the fast case code if we pass a number of guards. duke@435: // The paths which do not pass are accumulated in the slow_region. duke@435: RegionNode* slow_region = new (C, 1) RegionNode(1); duke@435: record_for_igvn(slow_region); duke@435: if (!stopped()) { duke@435: // It's an instance. Make the slow-path tests. duke@435: // If this is a virtual call, we generate a funny guard. We grab duke@435: // the vtable entry corresponding to clone() from the target object. duke@435: // If the target method which we are calling happens to be the duke@435: // Object clone() method, we pass the guard. We do not need this duke@435: // guard for non-virtual calls; the caller is known to be the native duke@435: // Object clone(). duke@435: if (is_virtual) { duke@435: generate_virtual_guard(obj_klass, slow_region); duke@435: } duke@435: duke@435: // The object must be cloneable and must not have a finalizer. duke@435: // Both of these conditions may be checked in a single test. duke@435: // We could optimize the cloneable test further, but we don't care. duke@435: generate_access_flags_guard(obj_klass, duke@435: // Test both conditions: duke@435: JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, duke@435: // Must be cloneable but not finalizer: duke@435: JVM_ACC_IS_CLONEABLE, duke@435: slow_region); duke@435: } duke@435: duke@435: if (!stopped()) { duke@435: // It's an instance, and it passed the slow-path tests. duke@435: PreserveJVMState pjvms(this); duke@435: Node* obj_size = NULL; duke@435: Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size); duke@435: assert(obj_size != NULL, ""); duke@435: Node* raw_obj = alloc_obj->in(1); duke@435: assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); duke@435: if (ReduceBulkZeroing) { duke@435: AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); duke@435: if (alloc != NULL && !alloc->maybe_set_complete(&_gvn)) duke@435: alloc = NULL; duke@435: } duke@435: if (!use_ReduceInitialCardMarks()) { duke@435: // Put in store barrier for any and all oops we are sticking duke@435: // into this object. (We could avoid this if we could prove duke@435: // that the object type contains no oop fields at all.) duke@435: card_mark = true; duke@435: } duke@435: alloc_val->init_req(_instance_alloc, raw_obj); duke@435: alloc_siz->init_req(_instance_alloc, obj_size); duke@435: alloc_reg->init_req(_instance_alloc, control()); duke@435: alloc_i_o->init_req(_instance_alloc, i_o()); duke@435: alloc_mem->init_req(_instance_alloc, memory(raw_adr_type)); duke@435: } duke@435: duke@435: // Generate code for the slow case. We make a call to clone(). duke@435: set_control(_gvn.transform(slow_region)); duke@435: if (!stopped()) { duke@435: PreserveJVMState pjvms(this); duke@435: CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); duke@435: Node* slow_result = set_results_for_java_call(slow_call); duke@435: // this->control() comes from set_results_for_java_call duke@435: result_reg->init_req(_slow_path, control()); duke@435: result_val->init_req(_slow_path, slow_result); duke@435: result_i_o ->set_req(_slow_path, i_o()); duke@435: result_mem ->set_req(_slow_path, reset_memory()); duke@435: } duke@435: duke@435: // The object is allocated, as an array and/or an instance. Now copy it. duke@435: set_control( _gvn.transform(alloc_reg) ); duke@435: set_i_o( _gvn.transform(alloc_i_o) ); duke@435: set_memory( _gvn.transform(alloc_mem), raw_adr_type ); duke@435: Node* raw_obj = _gvn.transform(alloc_val); duke@435: duke@435: if (!stopped()) { duke@435: // Copy the fastest available way. duke@435: // (No need for PreserveJVMState, since we're using it all up now.) kvn@598: // TODO: generate fields/elements copies for small objects instead. duke@435: Node* src = obj; duke@435: Node* dest = raw_obj; duke@435: Node* size = _gvn.transform(alloc_siz); duke@435: duke@435: // Exclude the header. coleenp@548: int base_off = instanceOopDesc::base_offset_in_bytes(); coleenp@548: if (UseCompressedOops) { kvn@598: assert(base_off % BytesPerLong != 0, "base with compressed oops"); kvn@598: // With compressed oops base_offset_in_bytes is 12 which creates kvn@598: // the gap since countx is rounded by 8 bytes below. kvn@598: // Copy klass and the gap. kvn@598: base_off = instanceOopDesc::klass_offset_in_bytes(); coleenp@548: } duke@435: src = basic_plus_adr(src, base_off); duke@435: dest = basic_plus_adr(dest, base_off); duke@435: duke@435: // Compute the length also, if needed: duke@435: Node* countx = size; duke@435: countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) ); duke@435: countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) )); duke@435: duke@435: // Select an appropriate instruction to initialize the range. duke@435: // The CopyArray instruction (if supported) can be optimized duke@435: // into a discrete set of scalar loads and stores. duke@435: bool disjoint_bases = true; duke@435: generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, duke@435: src, NULL, dest, NULL, countx); duke@435: duke@435: // Now that the object is properly initialized, type it as an oop. duke@435: // Use a secondary InitializeNode memory barrier. duke@435: InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, raw_adr_idx, duke@435: raw_obj)->as_Initialize(); duke@435: init->set_complete(&_gvn); // (there is no corresponding AllocateNode) duke@435: Node* new_obj = new(C, 2) CheckCastPPNode(control(), raw_obj, duke@435: TypeInstPtr::NOTNULL); duke@435: new_obj = _gvn.transform(new_obj); duke@435: duke@435: // If necessary, emit some card marks afterwards. (Non-arrays only.) duke@435: if (card_mark) { duke@435: Node* no_particular_value = NULL; duke@435: Node* no_particular_field = NULL; duke@435: post_barrier(control(), duke@435: memory(raw_adr_type), duke@435: new_obj, duke@435: no_particular_field, duke@435: raw_adr_idx, duke@435: no_particular_value, duke@435: T_OBJECT, duke@435: false); duke@435: } duke@435: // Present the results of the slow call. duke@435: result_reg->init_req(_fast_path, control()); duke@435: result_val->init_req(_fast_path, new_obj); duke@435: result_i_o ->set_req(_fast_path, i_o()); duke@435: result_mem ->set_req(_fast_path, reset_memory()); duke@435: } duke@435: duke@435: // Return the combined state. duke@435: set_control( _gvn.transform(result_reg) ); duke@435: set_i_o( _gvn.transform(result_i_o) ); duke@435: set_all_memory( _gvn.transform(result_mem) ); duke@435: duke@435: // Cast the result to a sharper type, since we know what clone does. duke@435: Node* new_obj = _gvn.transform(result_val); duke@435: Node* cast = new (C, 2) CheckCastPPNode(control(), new_obj, toop); duke@435: push(_gvn.transform(cast)); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: duke@435: // constants for computing the copy function duke@435: enum { duke@435: COPYFUNC_UNALIGNED = 0, duke@435: COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize duke@435: COPYFUNC_CONJOINT = 0, duke@435: COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend duke@435: }; duke@435: duke@435: // Note: The condition "disjoint" applies also for overlapping copies duke@435: // where an descending copy is permitted (i.e., dest_offset <= src_offset). duke@435: static address duke@435: select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) { duke@435: int selector = duke@435: (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) + duke@435: (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT); duke@435: duke@435: #define RETURN_STUB(xxx_arraycopy) { \ duke@435: name = #xxx_arraycopy; \ duke@435: return StubRoutines::xxx_arraycopy(); } duke@435: duke@435: switch (t) { duke@435: case T_BYTE: duke@435: case T_BOOLEAN: duke@435: switch (selector) { duke@435: case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy); duke@435: case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy); duke@435: case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy); duke@435: case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy); duke@435: } duke@435: case T_CHAR: duke@435: case T_SHORT: duke@435: switch (selector) { duke@435: case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy); duke@435: case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy); duke@435: case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy); duke@435: case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy); duke@435: } duke@435: case T_INT: duke@435: case T_FLOAT: duke@435: switch (selector) { duke@435: case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy); duke@435: case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy); duke@435: case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy); duke@435: case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy); duke@435: } duke@435: case T_DOUBLE: duke@435: case T_LONG: duke@435: switch (selector) { duke@435: case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy); duke@435: case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy); duke@435: case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy); duke@435: case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy); duke@435: } duke@435: case T_ARRAY: duke@435: case T_OBJECT: duke@435: switch (selector) { duke@435: case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy); duke@435: case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy); duke@435: case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy); duke@435: case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy); duke@435: } duke@435: default: duke@435: ShouldNotReachHere(); duke@435: return NULL; duke@435: } duke@435: duke@435: #undef RETURN_STUB duke@435: } duke@435: duke@435: //------------------------------basictype2arraycopy---------------------------- duke@435: address LibraryCallKit::basictype2arraycopy(BasicType t, duke@435: Node* src_offset, duke@435: Node* dest_offset, duke@435: bool disjoint_bases, duke@435: const char* &name) { duke@435: const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; duke@435: const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; duke@435: duke@435: bool aligned = false; duke@435: bool disjoint = disjoint_bases; duke@435: duke@435: // if the offsets are the same, we can treat the memory regions as duke@435: // disjoint, because either the memory regions are in different arrays, duke@435: // or they are identical (which we can treat as disjoint.) We can also duke@435: // treat a copy with a destination index less that the source index duke@435: // as disjoint since a low->high copy will work correctly in this case. duke@435: if (src_offset_inttype != NULL && src_offset_inttype->is_con() && duke@435: dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { duke@435: // both indices are constants duke@435: int s_offs = src_offset_inttype->get_con(); duke@435: int d_offs = dest_offset_inttype->get_con(); kvn@464: int element_size = type2aelembytes(t); duke@435: aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && duke@435: ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); duke@435: if (s_offs >= d_offs) disjoint = true; duke@435: } else if (src_offset == dest_offset && src_offset != NULL) { duke@435: // This can occur if the offsets are identical non-constants. duke@435: disjoint = true; duke@435: } duke@435: duke@435: return select_arraycopy_function(t, aligned, disjoint, name); duke@435: } duke@435: duke@435: duke@435: //------------------------------inline_arraycopy----------------------- duke@435: bool LibraryCallKit::inline_arraycopy() { duke@435: // Restore the stack and pop off the arguments. duke@435: int nargs = 5; // 2 oops, 3 ints, no size_t or long duke@435: assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); duke@435: duke@435: Node *src = argument(0); duke@435: Node *src_offset = argument(1); duke@435: Node *dest = argument(2); duke@435: Node *dest_offset = argument(3); duke@435: Node *length = argument(4); duke@435: duke@435: // Compile time checks. If any of these checks cannot be verified at compile time, duke@435: // we do not make a fast path for this call. Instead, we let the call remain as it duke@435: // is. The checks we choose to mandate at compile time are: duke@435: // duke@435: // (1) src and dest are arrays. duke@435: const Type* src_type = src->Value(&_gvn); duke@435: const Type* dest_type = dest->Value(&_gvn); duke@435: const TypeAryPtr* top_src = src_type->isa_aryptr(); duke@435: const TypeAryPtr* top_dest = dest_type->isa_aryptr(); duke@435: if (top_src == NULL || top_src->klass() == NULL || duke@435: top_dest == NULL || top_dest->klass() == NULL) { duke@435: // Conservatively insert a memory barrier on all memory slices. duke@435: // Do not let writes into the source float below the arraycopy. duke@435: insert_mem_bar(Op_MemBarCPUOrder); duke@435: duke@435: // Call StubRoutines::generic_arraycopy stub. duke@435: generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, duke@435: src, src_offset, dest, dest_offset, length, duke@435: nargs); duke@435: duke@435: // Do not let reads from the destination float above the arraycopy. duke@435: // Since we cannot type the arrays, we don't know which slices duke@435: // might be affected. We could restrict this barrier only to those duke@435: // memory slices which pertain to array elements--but don't bother. duke@435: if (!InsertMemBarAfterArraycopy) duke@435: // (If InsertMemBarAfterArraycopy, there is already one in place.) duke@435: insert_mem_bar(Op_MemBarCPUOrder); duke@435: return true; duke@435: } duke@435: duke@435: // (2) src and dest arrays must have elements of the same BasicType duke@435: // Figure out the size and type of the elements we will be copying. duke@435: BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); duke@435: BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); duke@435: if (src_elem == T_ARRAY) src_elem = T_OBJECT; duke@435: if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; duke@435: duke@435: if (src_elem != dest_elem || dest_elem == T_VOID) { duke@435: // The component types are not the same or are not recognized. Punt. duke@435: // (But, avoid the native method wrapper to JVM_ArrayCopy.) duke@435: generate_slow_arraycopy(TypePtr::BOTTOM, duke@435: src, src_offset, dest, dest_offset, length, duke@435: nargs); duke@435: return true; duke@435: } duke@435: duke@435: //--------------------------------------------------------------------------- duke@435: // We will make a fast path for this call to arraycopy. duke@435: duke@435: // We have the following tests left to perform: duke@435: // duke@435: // (3) src and dest must not be null. duke@435: // (4) src_offset must not be negative. duke@435: // (5) dest_offset must not be negative. duke@435: // (6) length must not be negative. duke@435: // (7) src_offset + length must not exceed length of src. duke@435: // (8) dest_offset + length must not exceed length of dest. duke@435: // (9) each element of an oop array must be assignable duke@435: duke@435: RegionNode* slow_region = new (C, 1) RegionNode(1); duke@435: record_for_igvn(slow_region); duke@435: duke@435: // (3) operands must not be null duke@435: // We currently perform our null checks with the do_null_check routine. duke@435: // This means that the null exceptions will be reported in the caller duke@435: // rather than (correctly) reported inside of the native arraycopy call. duke@435: // This should be corrected, given time. We do our null check with the duke@435: // stack pointer restored. duke@435: _sp += nargs; duke@435: src = do_null_check(src, T_ARRAY); duke@435: dest = do_null_check(dest, T_ARRAY); duke@435: _sp -= nargs; duke@435: duke@435: // (4) src_offset must not be negative. duke@435: generate_negative_guard(src_offset, slow_region); duke@435: duke@435: // (5) dest_offset must not be negative. duke@435: generate_negative_guard(dest_offset, slow_region); duke@435: duke@435: // (6) length must not be negative (moved to generate_arraycopy()). duke@435: // generate_negative_guard(length, slow_region); duke@435: duke@435: // (7) src_offset + length must not exceed length of src. duke@435: generate_limit_guard(src_offset, length, duke@435: load_array_length(src), duke@435: slow_region); duke@435: duke@435: // (8) dest_offset + length must not exceed length of dest. duke@435: generate_limit_guard(dest_offset, length, duke@435: load_array_length(dest), duke@435: slow_region); duke@435: duke@435: // (9) each element of an oop array must be assignable duke@435: // The generate_arraycopy subroutine checks this. duke@435: duke@435: // This is where the memory effects are placed: duke@435: const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); duke@435: generate_arraycopy(adr_type, dest_elem, duke@435: src, src_offset, dest, dest_offset, length, duke@435: nargs, false, false, slow_region); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: //-----------------------------generate_arraycopy---------------------- duke@435: // Generate an optimized call to arraycopy. duke@435: // Caller must guard against non-arrays. duke@435: // Caller must determine a common array basic-type for both arrays. duke@435: // Caller must validate offsets against array bounds. duke@435: // The slow_region has already collected guard failure paths duke@435: // (such as out of bounds length or non-conformable array types). duke@435: // The generated code has this shape, in general: duke@435: // duke@435: // if (length == 0) return // via zero_path duke@435: // slowval = -1 duke@435: // if (types unknown) { duke@435: // slowval = call generic copy loop duke@435: // if (slowval == 0) return // via checked_path duke@435: // } else if (indexes in bounds) { duke@435: // if ((is object array) && !(array type check)) { duke@435: // slowval = call checked copy loop duke@435: // if (slowval == 0) return // via checked_path duke@435: // } else { duke@435: // call bulk copy loop duke@435: // return // via fast_path duke@435: // } duke@435: // } duke@435: // // adjust params for remaining work: duke@435: // if (slowval != -1) { duke@435: // n = -1^slowval; src_offset += n; dest_offset += n; length -= n duke@435: // } duke@435: // slow_region: duke@435: // call slow arraycopy(src, src_offset, dest, dest_offset, length) duke@435: // return // via slow_call_path duke@435: // duke@435: // This routine is used from several intrinsics: System.arraycopy, duke@435: // Object.clone (the array subcase), and Arrays.copyOf[Range]. duke@435: // duke@435: void duke@435: LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, duke@435: BasicType basic_elem_type, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* copy_length, duke@435: int nargs, duke@435: bool disjoint_bases, duke@435: bool length_never_negative, duke@435: RegionNode* slow_region) { duke@435: duke@435: if (slow_region == NULL) { duke@435: slow_region = new(C,1) RegionNode(1); duke@435: record_for_igvn(slow_region); duke@435: } duke@435: duke@435: Node* original_dest = dest; duke@435: AllocateArrayNode* alloc = NULL; // used for zeroing, if needed duke@435: Node* raw_dest = NULL; // used before zeroing, if needed duke@435: bool must_clear_dest = false; duke@435: duke@435: // See if this is the initialization of a newly-allocated array. duke@435: // If so, we will take responsibility here for initializing it to zero. duke@435: // (Note: Because tightly_coupled_allocation performs checks on the duke@435: // out-edges of the dest, we need to avoid making derived pointers duke@435: // from it until we have checked its uses.) duke@435: if (ReduceBulkZeroing duke@435: && !ZeroTLAB // pointless if already zeroed duke@435: && basic_elem_type != T_CONFLICT // avoid corner case duke@435: && !_gvn.eqv_uncast(src, dest) duke@435: && ((alloc = tightly_coupled_allocation(dest, slow_region)) duke@435: != NULL) kvn@469: && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 duke@435: && alloc->maybe_set_complete(&_gvn)) { duke@435: // "You break it, you buy it." duke@435: InitializeNode* init = alloc->initialization(); duke@435: assert(init->is_complete(), "we just did this"); duke@435: assert(dest->Opcode() == Op_CheckCastPP, "sanity"); duke@435: assert(dest->in(0)->in(0) == init, "dest pinned"); duke@435: raw_dest = dest->in(1); // grab the raw pointer! duke@435: original_dest = dest; duke@435: dest = raw_dest; duke@435: adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory duke@435: // Decouple the original InitializeNode, turning it into a simple membar. duke@435: // We will build a new one at the end of this routine. duke@435: init->set_req(InitializeNode::RawAddress, top()); duke@435: // From this point on, every exit path is responsible for duke@435: // initializing any non-copied parts of the object to zero. duke@435: must_clear_dest = true; duke@435: } else { duke@435: // No zeroing elimination here. duke@435: alloc = NULL; duke@435: //original_dest = dest; duke@435: //must_clear_dest = false; duke@435: } duke@435: duke@435: // Results are placed here: duke@435: enum { fast_path = 1, // normal void-returning assembly stub duke@435: checked_path = 2, // special assembly stub with cleanup duke@435: slow_call_path = 3, // something went wrong; call the VM duke@435: zero_path = 4, // bypass when length of copy is zero duke@435: bcopy_path = 5, // copy primitive array by 64-bit blocks duke@435: PATH_LIMIT = 6 duke@435: }; duke@435: RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); duke@435: PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO); duke@435: PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type); duke@435: record_for_igvn(result_region); duke@435: _gvn.set_type_bottom(result_i_o); duke@435: _gvn.set_type_bottom(result_memory); duke@435: assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); duke@435: duke@435: // The slow_control path: duke@435: Node* slow_control; duke@435: Node* slow_i_o = i_o(); duke@435: Node* slow_mem = memory(adr_type); duke@435: debug_only(slow_control = (Node*) badAddress); duke@435: duke@435: // Checked control path: duke@435: Node* checked_control = top(); duke@435: Node* checked_mem = NULL; duke@435: Node* checked_i_o = NULL; duke@435: Node* checked_value = NULL; duke@435: duke@435: if (basic_elem_type == T_CONFLICT) { duke@435: assert(!must_clear_dest, ""); duke@435: Node* cv = generate_generic_arraycopy(adr_type, duke@435: src, src_offset, dest, dest_offset, duke@435: copy_length, nargs); duke@435: if (cv == NULL) cv = intcon(-1); // failure (no stub available) duke@435: checked_control = control(); duke@435: checked_i_o = i_o(); duke@435: checked_mem = memory(adr_type); duke@435: checked_value = cv; duke@435: set_control(top()); // no fast path duke@435: } duke@435: duke@435: Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); duke@435: if (not_pos != NULL) { duke@435: PreserveJVMState pjvms(this); duke@435: set_control(not_pos); duke@435: duke@435: // (6) length must not be negative. duke@435: if (!length_never_negative) { duke@435: generate_negative_guard(copy_length, slow_region); duke@435: } duke@435: duke@435: if (!stopped() && must_clear_dest) { duke@435: Node* dest_length = alloc->in(AllocateNode::ALength); duke@435: if (_gvn.eqv_uncast(copy_length, dest_length) duke@435: || _gvn.find_int_con(dest_length, 1) <= 0) { duke@435: // There is no zeroing to do. duke@435: } else { duke@435: // Clear the whole thing since there are no source elements to copy. duke@435: generate_clear_array(adr_type, dest, basic_elem_type, duke@435: intcon(0), NULL, duke@435: alloc->in(AllocateNode::AllocSize)); duke@435: } duke@435: } duke@435: duke@435: // Present the results of the fast call. duke@435: result_region->init_req(zero_path, control()); duke@435: result_i_o ->init_req(zero_path, i_o()); duke@435: result_memory->init_req(zero_path, memory(adr_type)); duke@435: } duke@435: duke@435: if (!stopped() && must_clear_dest) { duke@435: // We have to initialize the *uncopied* part of the array to zero. duke@435: // The copy destination is the slice dest[off..off+len]. The other slices duke@435: // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. duke@435: Node* dest_size = alloc->in(AllocateNode::AllocSize); duke@435: Node* dest_length = alloc->in(AllocateNode::ALength); duke@435: Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset, duke@435: copy_length) ); duke@435: duke@435: // If there is a head section that needs zeroing, do it now. duke@435: if (find_int_con(dest_offset, -1) != 0) { duke@435: generate_clear_array(adr_type, dest, basic_elem_type, duke@435: intcon(0), dest_offset, duke@435: NULL); duke@435: } duke@435: duke@435: // Next, perform a dynamic check on the tail length. duke@435: // It is often zero, and we can win big if we prove this. duke@435: // There are two wins: Avoid generating the ClearArray duke@435: // with its attendant messy index arithmetic, and upgrade duke@435: // the copy to a more hardware-friendly word size of 64 bits. duke@435: Node* tail_ctl = NULL; duke@435: if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) { duke@435: Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) ); duke@435: Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) ); duke@435: tail_ctl = generate_slow_guard(bol_lt, NULL); duke@435: assert(tail_ctl != NULL || !stopped(), "must be an outcome"); duke@435: } duke@435: duke@435: // At this point, let's assume there is no tail. duke@435: if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { duke@435: // There is no tail. Try an upgrade to a 64-bit copy. duke@435: bool didit = false; duke@435: { PreserveJVMState pjvms(this); duke@435: didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, duke@435: src, src_offset, dest, dest_offset, duke@435: dest_size); duke@435: if (didit) { duke@435: // Present the results of the block-copying fast call. duke@435: result_region->init_req(bcopy_path, control()); duke@435: result_i_o ->init_req(bcopy_path, i_o()); duke@435: result_memory->init_req(bcopy_path, memory(adr_type)); duke@435: } duke@435: } duke@435: if (didit) duke@435: set_control(top()); // no regular fast path duke@435: } duke@435: duke@435: // Clear the tail, if any. duke@435: if (tail_ctl != NULL) { duke@435: Node* notail_ctl = stopped() ? NULL : control(); duke@435: set_control(tail_ctl); duke@435: if (notail_ctl == NULL) { duke@435: generate_clear_array(adr_type, dest, basic_elem_type, duke@435: dest_tail, NULL, duke@435: dest_size); duke@435: } else { duke@435: // Make a local merge. duke@435: Node* done_ctl = new(C,3) RegionNode(3); duke@435: Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type); duke@435: done_ctl->init_req(1, notail_ctl); duke@435: done_mem->init_req(1, memory(adr_type)); duke@435: generate_clear_array(adr_type, dest, basic_elem_type, duke@435: dest_tail, NULL, duke@435: dest_size); duke@435: done_ctl->init_req(2, control()); duke@435: done_mem->init_req(2, memory(adr_type)); duke@435: set_control( _gvn.transform(done_ctl) ); duke@435: set_memory( _gvn.transform(done_mem), adr_type ); duke@435: } duke@435: } duke@435: } duke@435: duke@435: BasicType copy_type = basic_elem_type; duke@435: assert(basic_elem_type != T_ARRAY, "caller must fix this"); duke@435: if (!stopped() && copy_type == T_OBJECT) { duke@435: // If src and dest have compatible element types, we can copy bits. duke@435: // Types S[] and D[] are compatible if D is a supertype of S. duke@435: // duke@435: // If they are not, we will use checked_oop_disjoint_arraycopy, duke@435: // which performs a fast optimistic per-oop check, and backs off duke@435: // further to JVM_ArrayCopy on the first per-oop check that fails. duke@435: // (Actually, we don't move raw bits only; the GC requires card marks.) duke@435: duke@435: // Get the klassOop for both src and dest duke@435: Node* src_klass = load_object_klass(src); duke@435: Node* dest_klass = load_object_klass(dest); duke@435: duke@435: // Generate the subtype check. duke@435: // This might fold up statically, or then again it might not. duke@435: // duke@435: // Non-static example: Copying List.elements to a new String[]. duke@435: // The backing store for a List is always an Object[], duke@435: // but its elements are always type String, if the generic types duke@435: // are correct at the source level. duke@435: // duke@435: // Test S[] against D[], not S against D, because (probably) duke@435: // the secondary supertype cache is less busy for S[] than S. duke@435: // This usually only matters when D is an interface. duke@435: Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); duke@435: // Plug failing path into checked_oop_disjoint_arraycopy duke@435: if (not_subtype_ctrl != top()) { duke@435: PreserveJVMState pjvms(this); duke@435: set_control(not_subtype_ctrl); duke@435: // (At this point we can assume disjoint_bases, since types differ.) duke@435: int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc); duke@435: Node* p1 = basic_plus_adr(dest_klass, ek_offset); kvn@599: Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); duke@435: Node* dest_elem_klass = _gvn.transform(n1); duke@435: Node* cv = generate_checkcast_arraycopy(adr_type, duke@435: dest_elem_klass, duke@435: src, src_offset, dest, dest_offset, duke@435: copy_length, duke@435: nargs); duke@435: if (cv == NULL) cv = intcon(-1); // failure (no stub available) duke@435: checked_control = control(); duke@435: checked_i_o = i_o(); duke@435: checked_mem = memory(adr_type); duke@435: checked_value = cv; duke@435: } duke@435: // At this point we know we do not need type checks on oop stores. duke@435: duke@435: // Let's see if we need card marks: duke@435: if (alloc != NULL && use_ReduceInitialCardMarks()) { duke@435: // If we do not need card marks, copy using the jint or jlong stub. coleenp@548: copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); kvn@464: assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), duke@435: "sizes agree"); duke@435: } duke@435: } duke@435: duke@435: if (!stopped()) { duke@435: // Generate the fast path, if possible. duke@435: PreserveJVMState pjvms(this); duke@435: generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, duke@435: src, src_offset, dest, dest_offset, duke@435: ConvI2X(copy_length)); duke@435: duke@435: // Present the results of the fast call. duke@435: result_region->init_req(fast_path, control()); duke@435: result_i_o ->init_req(fast_path, i_o()); duke@435: result_memory->init_req(fast_path, memory(adr_type)); duke@435: } duke@435: duke@435: // Here are all the slow paths up to this point, in one bundle: duke@435: slow_control = top(); duke@435: if (slow_region != NULL) duke@435: slow_control = _gvn.transform(slow_region); duke@435: debug_only(slow_region = (RegionNode*)badAddress); duke@435: duke@435: set_control(checked_control); duke@435: if (!stopped()) { duke@435: // Clean up after the checked call. duke@435: // The returned value is either 0 or -1^K, duke@435: // where K = number of partially transferred array elements. duke@435: Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) ); duke@435: Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); duke@435: IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); duke@435: duke@435: // If it is 0, we are done, so transfer to the end. duke@435: Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) ); duke@435: result_region->init_req(checked_path, checks_done); duke@435: result_i_o ->init_req(checked_path, checked_i_o); duke@435: result_memory->init_req(checked_path, checked_mem); duke@435: duke@435: // If it is not zero, merge into the slow call. duke@435: set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) )); duke@435: RegionNode* slow_reg2 = new(C, 3) RegionNode(3); duke@435: PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO); duke@435: PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type); duke@435: record_for_igvn(slow_reg2); duke@435: slow_reg2 ->init_req(1, slow_control); duke@435: slow_i_o2 ->init_req(1, slow_i_o); duke@435: slow_mem2 ->init_req(1, slow_mem); duke@435: slow_reg2 ->init_req(2, control()); duke@435: slow_i_o2 ->init_req(2, i_o()); duke@435: slow_mem2 ->init_req(2, memory(adr_type)); duke@435: duke@435: slow_control = _gvn.transform(slow_reg2); duke@435: slow_i_o = _gvn.transform(slow_i_o2); duke@435: slow_mem = _gvn.transform(slow_mem2); duke@435: duke@435: if (alloc != NULL) { duke@435: // We'll restart from the very beginning, after zeroing the whole thing. duke@435: // This can cause double writes, but that's OK since dest is brand new. duke@435: // So we ignore the low 31 bits of the value returned from the stub. duke@435: } else { duke@435: // We must continue the copy exactly where it failed, or else duke@435: // another thread might see the wrong number of writes to dest. duke@435: Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) ); duke@435: Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT); duke@435: slow_offset->init_req(1, intcon(0)); duke@435: slow_offset->init_req(2, checked_offset); duke@435: slow_offset = _gvn.transform(slow_offset); duke@435: duke@435: // Adjust the arguments by the conditionally incoming offset. duke@435: Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) ); duke@435: Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) ); duke@435: Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) ); duke@435: duke@435: // Tweak the node variables to adjust the code produced below: duke@435: src_offset = src_off_plus; duke@435: dest_offset = dest_off_plus; duke@435: copy_length = length_minus; duke@435: } duke@435: } duke@435: duke@435: set_control(slow_control); duke@435: if (!stopped()) { duke@435: // Generate the slow path, if needed. duke@435: PreserveJVMState pjvms(this); // replace_in_map may trash the map duke@435: duke@435: set_memory(slow_mem, adr_type); duke@435: set_i_o(slow_i_o); duke@435: duke@435: if (must_clear_dest) { duke@435: generate_clear_array(adr_type, dest, basic_elem_type, duke@435: intcon(0), NULL, duke@435: alloc->in(AllocateNode::AllocSize)); duke@435: } duke@435: duke@435: if (dest != original_dest) { duke@435: // Promote from rawptr to oop, so it looks right in the call's GC map. duke@435: dest = _gvn.transform( new(C,2) CheckCastPPNode(control(), dest, duke@435: TypeInstPtr::NOTNULL) ); duke@435: duke@435: // Edit the call's debug-info to avoid referring to original_dest. duke@435: // (The problem with original_dest is that it isn't ready until duke@435: // after the InitializeNode completes, but this stuff is before.) duke@435: // Substitute in the locally valid dest_oop. duke@435: replace_in_map(original_dest, dest); duke@435: } duke@435: duke@435: generate_slow_arraycopy(adr_type, duke@435: src, src_offset, dest, dest_offset, duke@435: copy_length, nargs); duke@435: duke@435: result_region->init_req(slow_call_path, control()); duke@435: result_i_o ->init_req(slow_call_path, i_o()); duke@435: result_memory->init_req(slow_call_path, memory(adr_type)); duke@435: } duke@435: duke@435: // Remove unused edges. duke@435: for (uint i = 1; i < result_region->req(); i++) { duke@435: if (result_region->in(i) == NULL) duke@435: result_region->init_req(i, top()); duke@435: } duke@435: duke@435: // Finished; return the combined state. duke@435: set_control( _gvn.transform(result_region) ); duke@435: set_i_o( _gvn.transform(result_i_o) ); duke@435: set_memory( _gvn.transform(result_memory), adr_type ); duke@435: duke@435: if (dest != original_dest) { duke@435: // Pin the "finished" array node after the arraycopy/zeroing operations. duke@435: // Use a secondary InitializeNode memory barrier. duke@435: InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, duke@435: Compile::AliasIdxRaw, duke@435: raw_dest)->as_Initialize(); duke@435: init->set_complete(&_gvn); // (there is no corresponding AllocateNode) duke@435: _gvn.hash_delete(original_dest); duke@435: original_dest->set_req(0, control()); duke@435: _gvn.hash_find_insert(original_dest); // put back into GVN table duke@435: } duke@435: duke@435: // The memory edges above are precise in order to model effects around twisti@1040: // array copies accurately to allow value numbering of field loads around duke@435: // arraycopy. Such field loads, both before and after, are common in Java duke@435: // collections and similar classes involving header/array data structures. duke@435: // duke@435: // But with low number of register or when some registers are used or killed duke@435: // by arraycopy calls it causes registers spilling on stack. See 6544710. duke@435: // The next memory barrier is added to avoid it. If the arraycopy can be duke@435: // optimized away (which it can, sometimes) then we can manually remove duke@435: // the membar also. duke@435: if (InsertMemBarAfterArraycopy) duke@435: insert_mem_bar(Op_MemBarCPUOrder); duke@435: } duke@435: duke@435: duke@435: // Helper function which determines if an arraycopy immediately follows duke@435: // an allocation, with no intervening tests or other escapes for the object. duke@435: AllocateArrayNode* duke@435: LibraryCallKit::tightly_coupled_allocation(Node* ptr, duke@435: RegionNode* slow_region) { duke@435: if (stopped()) return NULL; // no fast path duke@435: if (C->AliasLevel() == 0) return NULL; // no MergeMems around duke@435: duke@435: AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); duke@435: if (alloc == NULL) return NULL; duke@435: duke@435: Node* rawmem = memory(Compile::AliasIdxRaw); duke@435: // Is the allocation's memory state untouched? duke@435: if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { duke@435: // Bail out if there have been raw-memory effects since the allocation. duke@435: // (Example: There might have been a call or safepoint.) duke@435: return NULL; duke@435: } duke@435: rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); duke@435: if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { duke@435: return NULL; duke@435: } duke@435: duke@435: // There must be no unexpected observers of this allocation. duke@435: for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { duke@435: Node* obs = ptr->fast_out(i); duke@435: if (obs != this->map()) { duke@435: return NULL; duke@435: } duke@435: } duke@435: duke@435: // This arraycopy must unconditionally follow the allocation of the ptr. duke@435: Node* alloc_ctl = ptr->in(0); duke@435: assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); duke@435: duke@435: Node* ctl = control(); duke@435: while (ctl != alloc_ctl) { duke@435: // There may be guards which feed into the slow_region. duke@435: // Any other control flow means that we might not get a chance duke@435: // to finish initializing the allocated object. duke@435: if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { duke@435: IfNode* iff = ctl->in(0)->as_If(); duke@435: Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); duke@435: assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); duke@435: if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { duke@435: ctl = iff->in(0); // This test feeds the known slow_region. duke@435: continue; duke@435: } duke@435: // One more try: Various low-level checks bottom out in duke@435: // uncommon traps. If the debug-info of the trap omits duke@435: // any reference to the allocation, as we've already duke@435: // observed, then there can be no objection to the trap. duke@435: bool found_trap = false; duke@435: for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { duke@435: Node* obs = not_ctl->fast_out(j); duke@435: if (obs->in(0) == not_ctl && obs->is_Call() && duke@435: (obs->as_Call()->entry_point() == duke@435: SharedRuntime::uncommon_trap_blob()->instructions_begin())) { duke@435: found_trap = true; break; duke@435: } duke@435: } duke@435: if (found_trap) { duke@435: ctl = iff->in(0); // This test feeds a harmless uncommon trap. duke@435: continue; duke@435: } duke@435: } duke@435: return NULL; duke@435: } duke@435: duke@435: // If we get this far, we have an allocation which immediately duke@435: // precedes the arraycopy, and we can take over zeroing the new object. duke@435: // The arraycopy will finish the initialization, and provide duke@435: // a new control state to which we will anchor the destination pointer. duke@435: duke@435: return alloc; duke@435: } duke@435: duke@435: // Helper for initialization of arrays, creating a ClearArray. duke@435: // It writes zero bits in [start..end), within the body of an array object. duke@435: // The memory effects are all chained onto the 'adr_type' alias category. duke@435: // duke@435: // Since the object is otherwise uninitialized, we are free duke@435: // to put a little "slop" around the edges of the cleared area, duke@435: // as long as it does not go back into the array's header, duke@435: // or beyond the array end within the heap. duke@435: // duke@435: // The lower edge can be rounded down to the nearest jint and the duke@435: // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. duke@435: // duke@435: // Arguments: duke@435: // adr_type memory slice where writes are generated duke@435: // dest oop of the destination array duke@435: // basic_elem_type element type of the destination duke@435: // slice_idx array index of first element to store duke@435: // slice_len number of elements to store (or NULL) duke@435: // dest_size total size in bytes of the array object duke@435: // duke@435: // Exactly one of slice_len or dest_size must be non-NULL. duke@435: // If dest_size is non-NULL, zeroing extends to the end of the object. duke@435: // If slice_len is non-NULL, the slice_idx value must be a constant. duke@435: void duke@435: LibraryCallKit::generate_clear_array(const TypePtr* adr_type, duke@435: Node* dest, duke@435: BasicType basic_elem_type, duke@435: Node* slice_idx, duke@435: Node* slice_len, duke@435: Node* dest_size) { duke@435: // one or the other but not both of slice_len and dest_size: duke@435: assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); duke@435: if (slice_len == NULL) slice_len = top(); duke@435: if (dest_size == NULL) dest_size = top(); duke@435: duke@435: // operate on this memory slice: duke@435: Node* mem = memory(adr_type); // memory slice to operate on duke@435: duke@435: // scaling and rounding of indexes: kvn@464: int scale = exact_log2(type2aelembytes(basic_elem_type)); duke@435: int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); duke@435: int clear_low = (-1 << scale) & (BytesPerInt - 1); duke@435: int bump_bit = (-1 << scale) & BytesPerInt; duke@435: duke@435: // determine constant starts and ends duke@435: const intptr_t BIG_NEG = -128; duke@435: assert(BIG_NEG + 2*abase < 0, "neg enough"); duke@435: intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); duke@435: intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); duke@435: if (slice_len_con == 0) { duke@435: return; // nothing to do here duke@435: } duke@435: intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; duke@435: intptr_t end_con = find_intptr_t_con(dest_size, -1); duke@435: if (slice_idx_con >= 0 && slice_len_con >= 0) { duke@435: assert(end_con < 0, "not two cons"); duke@435: end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), duke@435: BytesPerLong); duke@435: } duke@435: duke@435: if (start_con >= 0 && end_con >= 0) { duke@435: // Constant start and end. Simple. duke@435: mem = ClearArrayNode::clear_memory(control(), mem, dest, duke@435: start_con, end_con, &_gvn); duke@435: } else if (start_con >= 0 && dest_size != top()) { duke@435: // Constant start, pre-rounded end after the tail of the array. duke@435: Node* end = dest_size; duke@435: mem = ClearArrayNode::clear_memory(control(), mem, dest, duke@435: start_con, end, &_gvn); duke@435: } else if (start_con >= 0 && slice_len != top()) { duke@435: // Constant start, non-constant end. End needs rounding up. duke@435: // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) duke@435: intptr_t end_base = abase + (slice_idx_con << scale); duke@435: int end_round = (-1 << scale) & (BytesPerLong - 1); duke@435: Node* end = ConvI2X(slice_len); duke@435: if (scale != 0) duke@435: end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) )); duke@435: end_base += end_round; duke@435: end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) ); duke@435: end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) ); duke@435: mem = ClearArrayNode::clear_memory(control(), mem, dest, duke@435: start_con, end, &_gvn); duke@435: } else if (start_con < 0 && dest_size != top()) { duke@435: // Non-constant start, pre-rounded end after the tail of the array. duke@435: // This is almost certainly a "round-to-end" operation. duke@435: Node* start = slice_idx; duke@435: start = ConvI2X(start); duke@435: if (scale != 0) duke@435: start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) )); duke@435: start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) ); duke@435: if ((bump_bit | clear_low) != 0) { duke@435: int to_clear = (bump_bit | clear_low); duke@435: // Align up mod 8, then store a jint zero unconditionally duke@435: // just before the mod-8 boundary. coleenp@548: if (((abase + bump_bit) & ~to_clear) - bump_bit coleenp@548: < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { coleenp@548: bump_bit = 0; coleenp@548: assert((abase & to_clear) == 0, "array base must be long-aligned"); coleenp@548: } else { coleenp@548: // Bump 'start' up to (or past) the next jint boundary: coleenp@548: start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) ); coleenp@548: assert((abase & clear_low) == 0, "array base must be int-aligned"); coleenp@548: } duke@435: // Round bumped 'start' down to jlong boundary in body of array. duke@435: start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) ); coleenp@548: if (bump_bit != 0) { coleenp@548: // Store a zero to the immediately preceding jint: coleenp@548: Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) ); coleenp@548: Node* p1 = basic_plus_adr(dest, x1); coleenp@548: mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); coleenp@548: mem = _gvn.transform(mem); coleenp@548: } duke@435: } duke@435: Node* end = dest_size; // pre-rounded duke@435: mem = ClearArrayNode::clear_memory(control(), mem, dest, duke@435: start, end, &_gvn); duke@435: } else { duke@435: // Non-constant start, unrounded non-constant end. duke@435: // (Nobody zeroes a random midsection of an array using this routine.) duke@435: ShouldNotReachHere(); // fix caller duke@435: } duke@435: duke@435: // Done. duke@435: set_memory(mem, adr_type); duke@435: } duke@435: duke@435: duke@435: bool duke@435: LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, duke@435: BasicType basic_elem_type, duke@435: AllocateNode* alloc, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* dest_size) { duke@435: // See if there is an advantage from block transfer. kvn@464: int scale = exact_log2(type2aelembytes(basic_elem_type)); duke@435: if (scale >= LogBytesPerLong) duke@435: return false; // it is already a block transfer duke@435: duke@435: // Look at the alignment of the starting offsets. duke@435: int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); duke@435: const intptr_t BIG_NEG = -128; duke@435: assert(BIG_NEG + 2*abase < 0, "neg enough"); duke@435: duke@435: intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale); duke@435: intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale); duke@435: if (src_off < 0 || dest_off < 0) duke@435: // At present, we can only understand constants. duke@435: return false; duke@435: duke@435: if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { duke@435: // Non-aligned; too bad. duke@435: // One more chance: Pick off an initial 32-bit word. duke@435: // This is a common case, since abase can be odd mod 8. duke@435: if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && duke@435: ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { duke@435: Node* sptr = basic_plus_adr(src, src_off); duke@435: Node* dptr = basic_plus_adr(dest, dest_off); duke@435: Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); duke@435: store_to_memory(control(), dptr, sval, T_INT, adr_type); duke@435: src_off += BytesPerInt; duke@435: dest_off += BytesPerInt; duke@435: } else { duke@435: return false; duke@435: } duke@435: } duke@435: assert(src_off % BytesPerLong == 0, ""); duke@435: assert(dest_off % BytesPerLong == 0, ""); duke@435: duke@435: // Do this copy by giant steps. duke@435: Node* sptr = basic_plus_adr(src, src_off); duke@435: Node* dptr = basic_plus_adr(dest, dest_off); duke@435: Node* countx = dest_size; duke@435: countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) ); duke@435: countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) ); duke@435: duke@435: bool disjoint_bases = true; // since alloc != NULL duke@435: generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, duke@435: sptr, NULL, dptr, NULL, countx); duke@435: duke@435: return true; duke@435: } duke@435: duke@435: duke@435: // Helper function; generates code for the slow case. duke@435: // We make a call to a runtime method which emulates the native method, duke@435: // but without the native wrapper overhead. duke@435: void duke@435: LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* copy_length, duke@435: int nargs) { duke@435: _sp += nargs; // any deopt will start just before call to enclosing method duke@435: Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, duke@435: OptoRuntime::slow_arraycopy_Type(), duke@435: OptoRuntime::slow_arraycopy_Java(), duke@435: "slow_arraycopy", adr_type, duke@435: src, src_offset, dest, dest_offset, duke@435: copy_length); duke@435: _sp -= nargs; duke@435: duke@435: // Handle exceptions thrown by this fellow: duke@435: make_slow_call_ex(call, env()->Throwable_klass(), false); duke@435: } duke@435: duke@435: // Helper function; generates code for cases requiring runtime checks. duke@435: Node* duke@435: LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, duke@435: Node* dest_elem_klass, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* copy_length, duke@435: int nargs) { duke@435: if (stopped()) return NULL; duke@435: duke@435: address copyfunc_addr = StubRoutines::checkcast_arraycopy(); duke@435: if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. duke@435: return NULL; duke@435: } duke@435: duke@435: // Pick out the parameters required to perform a store-check duke@435: // for the target array. This is an optimistic check. It will duke@435: // look in each non-null element's class, at the desired klass's duke@435: // super_check_offset, for the desired klass. duke@435: int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc); duke@435: Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); duke@435: Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM); duke@435: Node* check_offset = _gvn.transform(n3); duke@435: Node* check_value = dest_elem_klass; duke@435: duke@435: Node* src_start = array_element_address(src, src_offset, T_OBJECT); duke@435: Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); duke@435: duke@435: // (We know the arrays are never conjoint, because their types differ.) duke@435: Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, duke@435: OptoRuntime::checkcast_arraycopy_Type(), duke@435: copyfunc_addr, "checkcast_arraycopy", adr_type, duke@435: // five arguments, of which two are duke@435: // intptr_t (jlong in LP64) duke@435: src_start, dest_start, duke@435: copy_length XTOP, duke@435: check_offset XTOP, duke@435: check_value); duke@435: duke@435: return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); duke@435: } duke@435: duke@435: duke@435: // Helper function; generates code for cases requiring runtime checks. duke@435: Node* duke@435: LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* copy_length, duke@435: int nargs) { duke@435: if (stopped()) return NULL; duke@435: duke@435: address copyfunc_addr = StubRoutines::generic_arraycopy(); duke@435: if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. duke@435: return NULL; duke@435: } duke@435: duke@435: Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, duke@435: OptoRuntime::generic_arraycopy_Type(), duke@435: copyfunc_addr, "generic_arraycopy", adr_type, duke@435: src, src_offset, dest, dest_offset, copy_length); duke@435: duke@435: return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); duke@435: } duke@435: duke@435: // Helper function; generates the fast out-of-line call to an arraycopy stub. duke@435: void duke@435: LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, duke@435: BasicType basic_elem_type, duke@435: bool disjoint_bases, duke@435: Node* src, Node* src_offset, duke@435: Node* dest, Node* dest_offset, duke@435: Node* copy_length) { duke@435: if (stopped()) return; // nothing to do duke@435: duke@435: Node* src_start = src; duke@435: Node* dest_start = dest; duke@435: if (src_offset != NULL || dest_offset != NULL) { duke@435: assert(src_offset != NULL && dest_offset != NULL, ""); duke@435: src_start = array_element_address(src, src_offset, basic_elem_type); duke@435: dest_start = array_element_address(dest, dest_offset, basic_elem_type); duke@435: } duke@435: duke@435: // Figure out which arraycopy runtime method to call. duke@435: const char* copyfunc_name = "arraycopy"; duke@435: address copyfunc_addr = duke@435: basictype2arraycopy(basic_elem_type, src_offset, dest_offset, duke@435: disjoint_bases, copyfunc_name); duke@435: duke@435: // Call it. Note that the count_ix value is not scaled to a byte-size. duke@435: make_runtime_call(RC_LEAF|RC_NO_FP, duke@435: OptoRuntime::fast_arraycopy_Type(), duke@435: copyfunc_addr, copyfunc_name, adr_type, duke@435: src_start, dest_start, copy_length XTOP); duke@435: }