• file
• revisions
• annotate
• diff
• comparison
• raw

src/share/vm/opto/library_call.cpp@9148c65abefc (annotated)

src/share/vm/opto/library_call.cpp

Thu, 29 May 2008 16:22:09 -0700

author

rasbold

date

Thu, 29 May 2008 16:22:09 -0700

changeset 604

9148c65abefc

parent 599

c436414a719e

child 609

510f98a80563

permissions

-rw-r--r--

6695049: (coll) Create an x86 intrinsic for Arrays.equals
Summary: Intrinsify java/util/Arrays.equals(char[], char[])
Reviewed-by: kvn, never

duke@435	1	/*
duke@435	2	* Copyright 1999-2007 Sun Microsystems, Inc. All Rights Reserved.
duke@435	3	* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435	4	*
duke@435	5	* This code is free software; you can redistribute it and/or modify it
duke@435	6	* under the terms of the GNU General Public License version 2 only, as
duke@435	7	* published by the Free Software Foundation.
duke@435	8	*
duke@435	9	* This code is distributed in the hope that it will be useful, but WITHOUT
duke@435	10	* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435	11	* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@435	12	* version 2 for more details (a copy is included in the LICENSE file that
duke@435	13	* accompanied this code).
duke@435	14	*
duke@435	15	* You should have received a copy of the GNU General Public License version
duke@435	16	* 2 along with this work; if not, write to the Free Software Foundation,
duke@435	17	* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435	18	*
duke@435	19	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
duke@435	20	* CA 95054 USA or visit www.sun.com if you need additional information or
duke@435	21	* have any questions.
duke@435	22	*
duke@435	23	*/
duke@435	24
duke@435	25	#include "incls/_precompiled.incl"
duke@435	26	#include "incls/_library_call.cpp.incl"
duke@435	27
duke@435	28	class LibraryIntrinsic : public InlineCallGenerator {
duke@435	29	// Extend the set of intrinsics known to the runtime:
duke@435	30	public:
duke@435	31	private:
duke@435	32	bool _is_virtual;
duke@435	33	vmIntrinsics::ID _intrinsic_id;
duke@435	34
duke@435	35	public:
duke@435	36	LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
duke@435	37	: InlineCallGenerator(m),
duke@435	38	_is_virtual(is_virtual),
duke@435	39	_intrinsic_id(id)
duke@435	40	{
duke@435	41	}
duke@435	42	virtual bool is_intrinsic() const { return true; }
duke@435	43	virtual bool is_virtual() const { return _is_virtual; }
duke@435	44	virtual JVMState* generate(JVMState* jvms);
duke@435	45	vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
duke@435	46	};
duke@435	47
duke@435	48
duke@435	49	// Local helper class for LibraryIntrinsic:
duke@435	50	class LibraryCallKit : public GraphKit {
duke@435	51	private:
duke@435	52	LibraryIntrinsic* _intrinsic; // the library intrinsic being called
duke@435	53
duke@435	54	public:
duke@435	55	LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
duke@435	56	: GraphKit(caller),
duke@435	57	_intrinsic(intrinsic)
duke@435	58	{
duke@435	59	}
duke@435	60
duke@435	61	ciMethod* caller() const { return jvms()->method(); }
duke@435	62	int bci() const { return jvms()->bci(); }
duke@435	63	LibraryIntrinsic* intrinsic() const { return _intrinsic; }
duke@435	64	vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
duke@435	65	ciMethod* callee() const { return _intrinsic->method(); }
duke@435	66	ciSignature* signature() const { return callee()->signature(); }
duke@435	67	int arg_size() const { return callee()->arg_size(); }
duke@435	68
duke@435	69	bool try_to_inline();
duke@435	70
duke@435	71	// Helper functions to inline natives
duke@435	72	void push_result(RegionNode* region, PhiNode* value);
duke@435	73	Node* generate_guard(Node* test, RegionNode* region, float true_prob);
duke@435	74	Node* generate_slow_guard(Node* test, RegionNode* region);
duke@435	75	Node* generate_fair_guard(Node* test, RegionNode* region);
duke@435	76	Node* generate_negative_guard(Node* index, RegionNode* region,
duke@435	77	// resulting CastII of index:
duke@435	78	Node* *pos_index = NULL);
duke@435	79	Node* generate_nonpositive_guard(Node* index, bool never_negative,
duke@435	80	// resulting CastII of index:
duke@435	81	Node* *pos_index = NULL);
duke@435	82	Node* generate_limit_guard(Node* offset, Node* subseq_length,
duke@435	83	Node* array_length,
duke@435	84	RegionNode* region);
duke@435	85	Node* generate_current_thread(Node* &tls_output);
duke@435	86	address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
duke@435	87	bool disjoint_bases, const char* &name);
duke@435	88	Node* load_mirror_from_klass(Node* klass);
duke@435	89	Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
duke@435	90	int nargs,
duke@435	91	RegionNode* region, int null_path,
duke@435	92	int offset);
duke@435	93	Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
duke@435	94	RegionNode* region, int null_path) {
duke@435	95	int offset = java_lang_Class::klass_offset_in_bytes();
duke@435	96	return load_klass_from_mirror_common(mirror, never_see_null, nargs,
duke@435	97	region, null_path,
duke@435	98	offset);
duke@435	99	}
duke@435	100	Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
duke@435	101	int nargs,
duke@435	102	RegionNode* region, int null_path) {
duke@435	103	int offset = java_lang_Class::array_klass_offset_in_bytes();
duke@435	104	return load_klass_from_mirror_common(mirror, never_see_null, nargs,
duke@435	105	region, null_path,
duke@435	106	offset);
duke@435	107	}
duke@435	108	Node* generate_access_flags_guard(Node* kls,
duke@435	109	int modifier_mask, int modifier_bits,
duke@435	110	RegionNode* region);
duke@435	111	Node* generate_interface_guard(Node* kls, RegionNode* region);
duke@435	112	Node* generate_array_guard(Node* kls, RegionNode* region) {
duke@435	113	return generate_array_guard_common(kls, region, false, false);
duke@435	114	}
duke@435	115	Node* generate_non_array_guard(Node* kls, RegionNode* region) {
duke@435	116	return generate_array_guard_common(kls, region, false, true);
duke@435	117	}
duke@435	118	Node* generate_objArray_guard(Node* kls, RegionNode* region) {
duke@435	119	return generate_array_guard_common(kls, region, true, false);
duke@435	120	}
duke@435	121	Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
duke@435	122	return generate_array_guard_common(kls, region, true, true);
duke@435	123	}
duke@435	124	Node* generate_array_guard_common(Node* kls, RegionNode* region,
duke@435	125	bool obj_array, bool not_array);
duke@435	126	Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
duke@435	127	CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
duke@435	128	bool is_virtual = false, bool is_static = false);
duke@435	129	CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
duke@435	130	return generate_method_call(method_id, false, true);
duke@435	131	}
duke@435	132	CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
duke@435	133	return generate_method_call(method_id, true, false);
duke@435	134	}
duke@435	135
duke@435	136	bool inline_string_compareTo();
duke@435	137	bool inline_string_indexOf();
duke@435	138	Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
duke@435	139	Node* pop_math_arg();
duke@435	140	bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
duke@435	141	bool inline_math_native(vmIntrinsics::ID id);
duke@435	142	bool inline_trig(vmIntrinsics::ID id);
duke@435	143	bool inline_trans(vmIntrinsics::ID id);
duke@435	144	bool inline_abs(vmIntrinsics::ID id);
duke@435	145	bool inline_sqrt(vmIntrinsics::ID id);
duke@435	146	bool inline_pow(vmIntrinsics::ID id);
duke@435	147	bool inline_exp(vmIntrinsics::ID id);
duke@435	148	bool inline_min_max(vmIntrinsics::ID id);
duke@435	149	Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
duke@435	150	// This returns Type::AnyPtr, RawPtr, or OopPtr.
duke@435	151	int classify_unsafe_addr(Node* &base, Node* &offset);
duke@435	152	Node* make_unsafe_address(Node* base, Node* offset);
duke@435	153	bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
duke@435	154	bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
duke@435	155	bool inline_unsafe_allocate();
duke@435	156	bool inline_unsafe_copyMemory();
duke@435	157	bool inline_native_currentThread();
duke@435	158	bool inline_native_time_funcs(bool isNano);
duke@435	159	bool inline_native_isInterrupted();
duke@435	160	bool inline_native_Class_query(vmIntrinsics::ID id);
duke@435	161	bool inline_native_subtype_check();
duke@435	162
duke@435	163	bool inline_native_newArray();
duke@435	164	bool inline_native_getLength();
duke@435	165	bool inline_array_copyOf(bool is_copyOfRange);
rasbold@604	166	bool inline_array_equals();
duke@435	167	bool inline_native_clone(bool is_virtual);
duke@435	168	bool inline_native_Reflection_getCallerClass();
duke@435	169	bool inline_native_AtomicLong_get();
duke@435	170	bool inline_native_AtomicLong_attemptUpdate();
duke@435	171	bool is_method_invoke_or_aux_frame(JVMState* jvms);
duke@435	172	// Helper function for inlining native object hash method
duke@435	173	bool inline_native_hashcode(bool is_virtual, bool is_static);
duke@435	174	bool inline_native_getClass();
duke@435	175
duke@435	176	// Helper functions for inlining arraycopy
duke@435	177	bool inline_arraycopy();
duke@435	178	void generate_arraycopy(const TypePtr* adr_type,
duke@435	179	BasicType basic_elem_type,
duke@435	180	Node* src, Node* src_offset,
duke@435	181	Node* dest, Node* dest_offset,
duke@435	182	Node* copy_length,
duke@435	183	int nargs, // arguments on stack for debug info
duke@435	184	bool disjoint_bases = false,
duke@435	185	bool length_never_negative = false,
duke@435	186	RegionNode* slow_region = NULL);
duke@435	187	AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
duke@435	188	RegionNode* slow_region);
duke@435	189	void generate_clear_array(const TypePtr* adr_type,
duke@435	190	Node* dest,
duke@435	191	BasicType basic_elem_type,
duke@435	192	Node* slice_off,
duke@435	193	Node* slice_len,
duke@435	194	Node* slice_end);
duke@435	195	bool generate_block_arraycopy(const TypePtr* adr_type,
duke@435	196	BasicType basic_elem_type,
duke@435	197	AllocateNode* alloc,
duke@435	198	Node* src, Node* src_offset,
duke@435	199	Node* dest, Node* dest_offset,
duke@435	200	Node* dest_size);
duke@435	201	void generate_slow_arraycopy(const TypePtr* adr_type,
duke@435	202	Node* src, Node* src_offset,
duke@435	203	Node* dest, Node* dest_offset,
duke@435	204	Node* copy_length,
duke@435	205	int nargs);
duke@435	206	Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
duke@435	207	Node* dest_elem_klass,
duke@435	208	Node* src, Node* src_offset,
duke@435	209	Node* dest, Node* dest_offset,
duke@435	210	Node* copy_length, int nargs);
duke@435	211	Node* generate_generic_arraycopy(const TypePtr* adr_type,
duke@435	212	Node* src, Node* src_offset,
duke@435	213	Node* dest, Node* dest_offset,
duke@435	214	Node* copy_length, int nargs);
duke@435	215	void generate_unchecked_arraycopy(const TypePtr* adr_type,
duke@435	216	BasicType basic_elem_type,
duke@435	217	bool disjoint_bases,
duke@435	218	Node* src, Node* src_offset,
duke@435	219	Node* dest, Node* dest_offset,
duke@435	220	Node* copy_length);
duke@435	221	bool inline_unsafe_CAS(BasicType type);
duke@435	222	bool inline_unsafe_ordered_store(BasicType type);
duke@435	223	bool inline_fp_conversions(vmIntrinsics::ID id);
duke@435	224	bool inline_reverseBytes(vmIntrinsics::ID id);
duke@435	225	};
duke@435	226
duke@435	227
duke@435	228	//---------------------------make_vm_intrinsic----------------------------
duke@435	229	CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
duke@435	230	vmIntrinsics::ID id = m->intrinsic_id();
duke@435	231	assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
duke@435	232
duke@435	233	if (DisableIntrinsic[0] != '\0'
duke@435	234	&& strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
duke@435	235	// disabled by a user request on the command line:
duke@435	236	// example: -XX:DisableIntrinsic=_hashCode,_getClass
duke@435	237	return NULL;
duke@435	238	}
duke@435	239
duke@435	240	if (!m->is_loaded()) {
duke@435	241	// do not attempt to inline unloaded methods
duke@435	242	return NULL;
duke@435	243	}
duke@435	244
duke@435	245	// Only a few intrinsics implement a virtual dispatch.
duke@435	246	// They are expensive calls which are also frequently overridden.
duke@435	247	if (is_virtual) {
duke@435	248	switch (id) {
duke@435	249	case vmIntrinsics::_hashCode:
duke@435	250	case vmIntrinsics::_clone:
duke@435	251	// OK, Object.hashCode and Object.clone intrinsics come in both flavors
duke@435	252	break;
duke@435	253	default:
duke@435	254	return NULL;
duke@435	255	}
duke@435	256	}
duke@435	257
duke@435	258	// -XX:-InlineNatives disables nearly all intrinsics:
duke@435	259	if (!InlineNatives) {
duke@435	260	switch (id) {
duke@435	261	case vmIntrinsics::_indexOf:
duke@435	262	case vmIntrinsics::_compareTo:
rasbold@604	263	case vmIntrinsics::_equalsC:
duke@435	264	break; // InlineNatives does not control String.compareTo
duke@435	265	default:
duke@435	266	return NULL;
duke@435	267	}
duke@435	268	}
duke@435	269
duke@435	270	switch (id) {
duke@435	271	case vmIntrinsics::_compareTo:
duke@435	272	if (!SpecialStringCompareTo) return NULL;
duke@435	273	break;
duke@435	274	case vmIntrinsics::_indexOf:
duke@435	275	if (!SpecialStringIndexOf) return NULL;
duke@435	276	break;
rasbold@604	277	case vmIntrinsics::_equalsC:
rasbold@604	278	if (!SpecialArraysEquals) return NULL;
rasbold@604	279	break;
duke@435	280	case vmIntrinsics::_arraycopy:
duke@435	281	if (!InlineArrayCopy) return NULL;
duke@435	282	break;
duke@435	283	case vmIntrinsics::_copyMemory:
duke@435	284	if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
duke@435	285	if (!InlineArrayCopy) return NULL;
duke@435	286	break;
duke@435	287	case vmIntrinsics::_hashCode:
duke@435	288	if (!InlineObjectHash) return NULL;
duke@435	289	break;
duke@435	290	case vmIntrinsics::_clone:
duke@435	291	case vmIntrinsics::_copyOf:
duke@435	292	case vmIntrinsics::_copyOfRange:
duke@435	293	if (!InlineObjectCopy) return NULL;
duke@435	294	// These also use the arraycopy intrinsic mechanism:
duke@435	295	if (!InlineArrayCopy) return NULL;
duke@435	296	break;
duke@435	297	case vmIntrinsics::_checkIndex:
duke@435	298	// We do not intrinsify this. The optimizer does fine with it.
duke@435	299	return NULL;
duke@435	300
duke@435	301	case vmIntrinsics::_get_AtomicLong:
duke@435	302	case vmIntrinsics::_attemptUpdate:
duke@435	303	if (!InlineAtomicLong) return NULL;
duke@435	304	break;
duke@435	305
duke@435	306	case vmIntrinsics::_Object_init:
duke@435	307	case vmIntrinsics::_invoke:
duke@435	308	// We do not intrinsify these; they are marked for other purposes.
duke@435	309	return NULL;
duke@435	310
duke@435	311	case vmIntrinsics::_getCallerClass:
duke@435	312	if (!UseNewReflection) return NULL;
duke@435	313	if (!InlineReflectionGetCallerClass) return NULL;
duke@435	314	if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
duke@435	315	break;
duke@435	316
duke@435	317	default:
duke@435	318	break;
duke@435	319	}
duke@435	320
duke@435	321	// -XX:-InlineClassNatives disables natives from the Class class.
duke@435	322	// The flag applies to all reflective calls, notably Array.newArray
duke@435	323	// (visible to Java programmers as Array.newInstance).
duke@435	324	if (m->holder()->name() == ciSymbol::java_lang_Class() ||
duke@435	325	m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
duke@435	326	if (!InlineClassNatives) return NULL;
duke@435	327	}
duke@435	328
duke@435	329	// -XX:-InlineThreadNatives disables natives from the Thread class.
duke@435	330	if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
duke@435	331	if (!InlineThreadNatives) return NULL;
duke@435	332	}
duke@435	333
duke@435	334	// -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
duke@435	335	if (m->holder()->name() == ciSymbol::java_lang_Math() ||
duke@435	336	m->holder()->name() == ciSymbol::java_lang_Float() ||
duke@435	337	m->holder()->name() == ciSymbol::java_lang_Double()) {
duke@435	338	if (!InlineMathNatives) return NULL;
duke@435	339	}
duke@435	340
duke@435	341	// -XX:-InlineUnsafeOps disables natives from the Unsafe class.
duke@435	342	if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
duke@435	343	if (!InlineUnsafeOps) return NULL;
duke@435	344	}
duke@435	345
duke@435	346	return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
duke@435	347	}
duke@435	348
duke@435	349	//----------------------register_library_intrinsics-----------------------
duke@435	350	// Initialize this file's data structures, for each Compile instance.
duke@435	351	void Compile::register_library_intrinsics() {
duke@435	352	// Nothing to do here.
duke@435	353	}
duke@435	354
duke@435	355	JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
duke@435	356	LibraryCallKit kit(jvms, this);
duke@435	357	Compile* C = kit.C;
duke@435	358	int nodes = C->unique();
duke@435	359	#ifndef PRODUCT
duke@435	360	if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
duke@435	361	char buf[1000];
duke@435	362	const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
duke@435	363	tty->print_cr("Intrinsic %s", str);
duke@435	364	}
duke@435	365	#endif
duke@435	366	if (kit.try_to_inline()) {
duke@435	367	if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
duke@435	368	tty->print("Inlining intrinsic %s%s at bci:%d in",
duke@435	369	vmIntrinsics::name_at(intrinsic_id()),
duke@435	370	(is_virtual() ? " (virtual)" : ""), kit.bci());
duke@435	371	kit.caller()->print_short_name(tty);
duke@435	372	tty->print_cr(" (%d bytes)", kit.caller()->code_size());
duke@435	373	}
duke@435	374	C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
duke@435	375	if (C->log()) {
duke@435	376	C->log()->elem("intrinsic id='%s'%s nodes='%d'",
duke@435	377	vmIntrinsics::name_at(intrinsic_id()),
duke@435	378	(is_virtual() ? " virtual='1'" : ""),
duke@435	379	C->unique() - nodes);
duke@435	380	}
duke@435	381	return kit.transfer_exceptions_into_jvms();
duke@435	382	}
duke@435	383
duke@435	384	if (PrintIntrinsics) {
duke@435	385	switch (intrinsic_id()) {
duke@435	386	case vmIntrinsics::_invoke:
duke@435	387	case vmIntrinsics::_Object_init:
duke@435	388	// We do not expect to inline these, so do not produce any noise about them.
duke@435	389	break;
duke@435	390	default:
duke@435	391	tty->print("Did not inline intrinsic %s%s at bci:%d in",
duke@435	392	vmIntrinsics::name_at(intrinsic_id()),
duke@435	393	(is_virtual() ? " (virtual)" : ""), kit.bci());
duke@435	394	kit.caller()->print_short_name(tty);
duke@435	395	tty->print_cr(" (%d bytes)", kit.caller()->code_size());
duke@435	396	}
duke@435	397	}
duke@435	398	C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
duke@435	399	return NULL;
duke@435	400	}
duke@435	401
duke@435	402	bool LibraryCallKit::try_to_inline() {
duke@435	403	// Handle symbolic names for otherwise undistinguished boolean switches:
duke@435	404	const bool is_store = true;
duke@435	405	const bool is_native_ptr = true;
duke@435	406	const bool is_static = true;
duke@435	407
duke@435	408	switch (intrinsic_id()) {
duke@435	409	case vmIntrinsics::_hashCode:
duke@435	410	return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
duke@435	411	case vmIntrinsics::_identityHashCode:
duke@435	412	return inline_native_hashcode(/*!virtual*/ false, is_static);
duke@435	413	case vmIntrinsics::_getClass:
duke@435	414	return inline_native_getClass();
duke@435	415
duke@435	416	case vmIntrinsics::_dsin:
duke@435	417	case vmIntrinsics::_dcos:
duke@435	418	case vmIntrinsics::_dtan:
duke@435	419	case vmIntrinsics::_dabs:
duke@435	420	case vmIntrinsics::_datan2:
duke@435	421	case vmIntrinsics::_dsqrt:
duke@435	422	case vmIntrinsics::_dexp:
duke@435	423	case vmIntrinsics::_dlog:
duke@435	424	case vmIntrinsics::_dlog10:
duke@435	425	case vmIntrinsics::_dpow:
duke@435	426	return inline_math_native(intrinsic_id());
duke@435	427
duke@435	428	case vmIntrinsics::_min:
duke@435	429	case vmIntrinsics::_max:
duke@435	430	return inline_min_max(intrinsic_id());
duke@435	431
duke@435	432	case vmIntrinsics::_arraycopy:
duke@435	433	return inline_arraycopy();
duke@435	434
duke@435	435	case vmIntrinsics::_compareTo:
duke@435	436	return inline_string_compareTo();
duke@435	437	case vmIntrinsics::_indexOf:
duke@435	438	return inline_string_indexOf();
duke@435	439
duke@435	440	case vmIntrinsics::_getObject:
duke@435	441	return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
duke@435	442	case vmIntrinsics::_getBoolean:
duke@435	443	return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
duke@435	444	case vmIntrinsics::_getByte:
duke@435	445	return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
duke@435	446	case vmIntrinsics::_getShort:
duke@435	447	return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
duke@435	448	case vmIntrinsics::_getChar:
duke@435	449	return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
duke@435	450	case vmIntrinsics::_getInt:
duke@435	451	return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
duke@435	452	case vmIntrinsics::_getLong:
duke@435	453	return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
duke@435	454	case vmIntrinsics::_getFloat:
duke@435	455	return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
duke@435	456	case vmIntrinsics::_getDouble:
duke@435	457	return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
duke@435	458
duke@435	459	case vmIntrinsics::_putObject:
duke@435	460	return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
duke@435	461	case vmIntrinsics::_putBoolean:
duke@435	462	return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
duke@435	463	case vmIntrinsics::_putByte:
duke@435	464	return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
duke@435	465	case vmIntrinsics::_putShort:
duke@435	466	return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
duke@435	467	case vmIntrinsics::_putChar:
duke@435	468	return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
duke@435	469	case vmIntrinsics::_putInt:
duke@435	470	return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
duke@435	471	case vmIntrinsics::_putLong:
duke@435	472	return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
duke@435	473	case vmIntrinsics::_putFloat:
duke@435	474	return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
duke@435	475	case vmIntrinsics::_putDouble:
duke@435	476	return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
duke@435	477
duke@435	478	case vmIntrinsics::_getByte_raw:
duke@435	479	return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
duke@435	480	case vmIntrinsics::_getShort_raw:
duke@435	481	return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
duke@435	482	case vmIntrinsics::_getChar_raw:
duke@435	483	return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
duke@435	484	case vmIntrinsics::_getInt_raw:
duke@435	485	return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
duke@435	486	case vmIntrinsics::_getLong_raw:
duke@435	487	return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
duke@435	488	case vmIntrinsics::_getFloat_raw:
duke@435	489	return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
duke@435	490	case vmIntrinsics::_getDouble_raw:
duke@435	491	return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
duke@435	492	case vmIntrinsics::_getAddress_raw:
duke@435	493	return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
duke@435	494
duke@435	495	case vmIntrinsics::_putByte_raw:
duke@435	496	return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
duke@435	497	case vmIntrinsics::_putShort_raw:
duke@435	498	return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
duke@435	499	case vmIntrinsics::_putChar_raw:
duke@435	500	return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
duke@435	501	case vmIntrinsics::_putInt_raw:
duke@435	502	return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
duke@435	503	case vmIntrinsics::_putLong_raw:
duke@435	504	return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
duke@435	505	case vmIntrinsics::_putFloat_raw:
duke@435	506	return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
duke@435	507	case vmIntrinsics::_putDouble_raw:
duke@435	508	return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
duke@435	509	case vmIntrinsics::_putAddress_raw:
duke@435	510	return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
duke@435	511
duke@435	512	case vmIntrinsics::_getObjectVolatile:
duke@435	513	return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
duke@435	514	case vmIntrinsics::_getBooleanVolatile:
duke@435	515	return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
duke@435	516	case vmIntrinsics::_getByteVolatile:
duke@435	517	return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
duke@435	518	case vmIntrinsics::_getShortVolatile:
duke@435	519	return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
duke@435	520	case vmIntrinsics::_getCharVolatile:
duke@435	521	return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
duke@435	522	case vmIntrinsics::_getIntVolatile:
duke@435	523	return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
duke@435	524	case vmIntrinsics::_getLongVolatile:
duke@435	525	return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
duke@435	526	case vmIntrinsics::_getFloatVolatile:
duke@435	527	return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
duke@435	528	case vmIntrinsics::_getDoubleVolatile:
duke@435	529	return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
duke@435	530
duke@435	531	case vmIntrinsics::_putObjectVolatile:
duke@435	532	return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
duke@435	533	case vmIntrinsics::_putBooleanVolatile:
duke@435	534	return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
duke@435	535	case vmIntrinsics::_putByteVolatile:
duke@435	536	return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
duke@435	537	case vmIntrinsics::_putShortVolatile:
duke@435	538	return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
duke@435	539	case vmIntrinsics::_putCharVolatile:
duke@435	540	return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
duke@435	541	case vmIntrinsics::_putIntVolatile:
duke@435	542	return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
duke@435	543	case vmIntrinsics::_putLongVolatile:
duke@435	544	return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
duke@435	545	case vmIntrinsics::_putFloatVolatile:
duke@435	546	return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
duke@435	547	case vmIntrinsics::_putDoubleVolatile:
duke@435	548	return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
duke@435	549
duke@435	550	case vmIntrinsics::_prefetchRead:
duke@435	551	return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
duke@435	552	case vmIntrinsics::_prefetchWrite:
duke@435	553	return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
duke@435	554	case vmIntrinsics::_prefetchReadStatic:
duke@435	555	return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
duke@435	556	case vmIntrinsics::_prefetchWriteStatic:
duke@435	557	return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
duke@435	558
duke@435	559	case vmIntrinsics::_compareAndSwapObject:
duke@435	560	return inline_unsafe_CAS(T_OBJECT);
duke@435	561	case vmIntrinsics::_compareAndSwapInt:
duke@435	562	return inline_unsafe_CAS(T_INT);
duke@435	563	case vmIntrinsics::_compareAndSwapLong:
duke@435	564	return inline_unsafe_CAS(T_LONG);
duke@435	565
duke@435	566	case vmIntrinsics::_putOrderedObject:
duke@435	567	return inline_unsafe_ordered_store(T_OBJECT);
duke@435	568	case vmIntrinsics::_putOrderedInt:
duke@435	569	return inline_unsafe_ordered_store(T_INT);
duke@435	570	case vmIntrinsics::_putOrderedLong:
duke@435	571	return inline_unsafe_ordered_store(T_LONG);
duke@435	572
duke@435	573	case vmIntrinsics::_currentThread:
duke@435	574	return inline_native_currentThread();
duke@435	575	case vmIntrinsics::_isInterrupted:
duke@435	576	return inline_native_isInterrupted();
duke@435	577
duke@435	578	case vmIntrinsics::_currentTimeMillis:
duke@435	579	return inline_native_time_funcs(false);
duke@435	580	case vmIntrinsics::_nanoTime:
duke@435	581	return inline_native_time_funcs(true);
duke@435	582	case vmIntrinsics::_allocateInstance:
duke@435	583	return inline_unsafe_allocate();
duke@435	584	case vmIntrinsics::_copyMemory:
duke@435	585	return inline_unsafe_copyMemory();
duke@435	586	case vmIntrinsics::_newArray:
duke@435	587	return inline_native_newArray();
duke@435	588	case vmIntrinsics::_getLength:
duke@435	589	return inline_native_getLength();
duke@435	590	case vmIntrinsics::_copyOf:
duke@435	591	return inline_array_copyOf(false);
duke@435	592	case vmIntrinsics::_copyOfRange:
duke@435	593	return inline_array_copyOf(true);
rasbold@604	594	case vmIntrinsics::_equalsC:
rasbold@604	595	return inline_array_equals();
duke@435	596	case vmIntrinsics::_clone:
duke@435	597	return inline_native_clone(intrinsic()->is_virtual());
duke@435	598
duke@435	599	case vmIntrinsics::_isAssignableFrom:
duke@435	600	return inline_native_subtype_check();
duke@435	601
duke@435	602	case vmIntrinsics::_isInstance:
duke@435	603	case vmIntrinsics::_getModifiers:
duke@435	604	case vmIntrinsics::_isInterface:
duke@435	605	case vmIntrinsics::_isArray:
duke@435	606	case vmIntrinsics::_isPrimitive:
duke@435	607	case vmIntrinsics::_getSuperclass:
duke@435	608	case vmIntrinsics::_getComponentType:
duke@435	609	case vmIntrinsics::_getClassAccessFlags:
duke@435	610	return inline_native_Class_query(intrinsic_id());
duke@435	611
duke@435	612	case vmIntrinsics::_floatToRawIntBits:
duke@435	613	case vmIntrinsics::_floatToIntBits:
duke@435	614	case vmIntrinsics::_intBitsToFloat:
duke@435	615	case vmIntrinsics::_doubleToRawLongBits:
duke@435	616	case vmIntrinsics::_doubleToLongBits:
duke@435	617	case vmIntrinsics::_longBitsToDouble:
duke@435	618	return inline_fp_conversions(intrinsic_id());
duke@435	619
duke@435	620	case vmIntrinsics::_reverseBytes_i:
duke@435	621	case vmIntrinsics::_reverseBytes_l:
duke@435	622	return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
duke@435	623
duke@435	624	case vmIntrinsics::_get_AtomicLong:
duke@435	625	return inline_native_AtomicLong_get();
duke@435	626	case vmIntrinsics::_attemptUpdate:
duke@435	627	return inline_native_AtomicLong_attemptUpdate();
duke@435	628
duke@435	629	case vmIntrinsics::_getCallerClass:
duke@435	630	return inline_native_Reflection_getCallerClass();
duke@435	631
duke@435	632	default:
duke@435	633	// If you get here, it may be that someone has added a new intrinsic
duke@435	634	// to the list in vmSymbols.hpp without implementing it here.
duke@435	635	#ifndef PRODUCT
duke@435	636	if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
duke@435	637	tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
duke@435	638	vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
duke@435	639	}
duke@435	640	#endif
duke@435	641	return false;
duke@435	642	}
duke@435	643	}
duke@435	644
duke@435	645	//------------------------------push_result------------------------------
duke@435	646	// Helper function for finishing intrinsics.
duke@435	647	void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
duke@435	648	record_for_igvn(region);
duke@435	649	set_control(_gvn.transform(region));
duke@435	650	BasicType value_type = value->type()->basic_type();
duke@435	651	push_node(value_type, _gvn.transform(value));
duke@435	652	}
duke@435	653
duke@435	654	//------------------------------generate_guard---------------------------
duke@435	655	// Helper function for generating guarded fast-slow graph structures.
duke@435	656	// The given 'test', if true, guards a slow path. If the test fails
duke@435	657	// then a fast path can be taken. (We generally hope it fails.)
duke@435	658	// In all cases, GraphKit::control() is updated to the fast path.
duke@435	659	// The returned value represents the control for the slow path.
duke@435	660	// The return value is never 'top'; it is either a valid control
duke@435	661	// or NULL if it is obvious that the slow path can never be taken.
duke@435	662	// Also, if region and the slow control are not NULL, the slow edge
duke@435	663	// is appended to the region.
duke@435	664	Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
duke@435	665	if (stopped()) {
duke@435	666	// Already short circuited.
duke@435	667	return NULL;
duke@435	668	}
duke@435	669
duke@435	670	// Build an if node and its projections.
duke@435	671	// If test is true we take the slow path, which we assume is uncommon.
duke@435	672	if (_gvn.type(test) == TypeInt::ZERO) {
duke@435	673	// The slow branch is never taken. No need to build this guard.
duke@435	674	return NULL;
duke@435	675	}
duke@435	676
duke@435	677	IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
duke@435	678
duke@435	679	Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
duke@435	680	if (if_slow == top()) {
duke@435	681	// The slow branch is never taken. No need to build this guard.
duke@435	682	return NULL;
duke@435	683	}
duke@435	684
duke@435	685	if (region != NULL)
duke@435	686	region->add_req(if_slow);
duke@435	687
duke@435	688	Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
duke@435	689	set_control(if_fast);
duke@435	690
duke@435	691	return if_slow;
duke@435	692	}
duke@435	693
duke@435	694	inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
duke@435	695	return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
duke@435	696	}
duke@435	697	inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
duke@435	698	return generate_guard(test, region, PROB_FAIR);
duke@435	699	}
duke@435	700
duke@435	701	inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
duke@435	702	Node* *pos_index) {
duke@435	703	if (stopped())
duke@435	704	return NULL; // already stopped
duke@435	705	if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
duke@435	706	return NULL; // index is already adequately typed
duke@435	707	Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
duke@435	708	Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
duke@435	709	Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
duke@435	710	if (is_neg != NULL && pos_index != NULL) {
duke@435	711	// Emulate effect of Parse::adjust_map_after_if.
duke@435	712	Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
duke@435	713	ccast->set_req(0, control());
duke@435	714	(*pos_index) = _gvn.transform(ccast);
duke@435	715	}
duke@435	716	return is_neg;
duke@435	717	}
duke@435	718
duke@435	719	inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
duke@435	720	Node* *pos_index) {
duke@435	721	if (stopped())
duke@435	722	return NULL; // already stopped
duke@435	723	if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
duke@435	724	return NULL; // index is already adequately typed
duke@435	725	Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
duke@435	726	BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
duke@435	727	Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
duke@435	728	Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
duke@435	729	if (is_notp != NULL && pos_index != NULL) {
duke@435	730	// Emulate effect of Parse::adjust_map_after_if.
duke@435	731	Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
duke@435	732	ccast->set_req(0, control());
duke@435	733	(*pos_index) = _gvn.transform(ccast);
duke@435	734	}
duke@435	735	return is_notp;
duke@435	736	}
duke@435	737
duke@435	738	// Make sure that 'position' is a valid limit index, in [0..length].
duke@435	739	// There are two equivalent plans for checking this:
duke@435	740	// A. (offset + copyLength) unsigned<= arrayLength
duke@435	741	// B. offset <= (arrayLength - copyLength)
duke@435	742	// We require that all of the values above, except for the sum and
duke@435	743	// difference, are already known to be non-negative.
duke@435	744	// Plan A is robust in the face of overflow, if offset and copyLength
duke@435	745	// are both hugely positive.
duke@435	746	//
duke@435	747	// Plan B is less direct and intuitive, but it does not overflow at
duke@435	748	// all, since the difference of two non-negatives is always
duke@435	749	// representable. Whenever Java methods must perform the equivalent
duke@435	750	// check they generally use Plan B instead of Plan A.
duke@435	751	// For the moment we use Plan A.
duke@435	752	inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
duke@435	753	Node* subseq_length,
duke@435	754	Node* array_length,
duke@435	755	RegionNode* region) {
duke@435	756	if (stopped())
duke@435	757	return NULL; // already stopped
duke@435	758	bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
duke@435	759	if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
duke@435	760	return NULL; // common case of whole-array copy
duke@435	761	Node* last = subseq_length;
duke@435	762	if (!zero_offset) // last += offset
duke@435	763	last = _gvn.transform( new (C, 3) AddINode(last, offset));
duke@435	764	Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
duke@435	765	Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
duke@435	766	Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
duke@435	767	return is_over;
duke@435	768	}
duke@435	769
duke@435	770
duke@435	771	//--------------------------generate_current_thread--------------------
duke@435	772	Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
duke@435	773	ciKlass* thread_klass = env()->Thread_klass();
duke@435	774	const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
duke@435	775	Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
duke@435	776	Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
duke@435	777	Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
duke@435	778	tls_output = thread;
duke@435	779	return threadObj;
duke@435	780	}
duke@435	781
duke@435	782
duke@435	783	//------------------------------inline_string_compareTo------------------------
duke@435	784	bool LibraryCallKit::inline_string_compareTo() {
duke@435	785
duke@435	786	const int value_offset = java_lang_String::value_offset_in_bytes();
duke@435	787	const int count_offset = java_lang_String::count_offset_in_bytes();
duke@435	788	const int offset_offset = java_lang_String::offset_offset_in_bytes();
duke@435	789
duke@435	790	_sp += 2;
duke@435	791	Node *argument = pop(); // pop non-receiver first: it was pushed second
duke@435	792	Node *receiver = pop();
duke@435	793
duke@435	794	// Null check on self without removing any arguments. The argument
duke@435	795	// null check technically happens in the wrong place, which can lead to
duke@435	796	// invalid stack traces when string compare is inlined into a method
duke@435	797	// which handles NullPointerExceptions.
duke@435	798	_sp += 2;
duke@435	799	receiver = do_null_check(receiver, T_OBJECT);
duke@435	800	argument = do_null_check(argument, T_OBJECT);
duke@435	801	_sp -= 2;
duke@435	802	if (stopped()) {
duke@435	803	return true;
duke@435	804	}
duke@435	805
duke@435	806	ciInstanceKlass* klass = env()->String_klass();
duke@435	807	const TypeInstPtr* string_type =
duke@435	808	TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
duke@435	809
duke@435	810	Node* compare =
duke@435	811	_gvn.transform(new (C, 7) StrCompNode(
duke@435	812	control(),
duke@435	813	memory(TypeAryPtr::CHARS),
duke@435	814	memory(string_type->add_offset(value_offset)),
duke@435	815	memory(string_type->add_offset(count_offset)),
duke@435	816	memory(string_type->add_offset(offset_offset)),
duke@435	817	receiver,
duke@435	818	argument));
duke@435	819	push(compare);
duke@435	820	return true;
duke@435	821	}
duke@435	822
rasbold@604	823	//------------------------------inline_array_equals----------------------------
rasbold@604	824	bool LibraryCallKit::inline_array_equals() {
rasbold@604	825
rasbold@604	826	_sp += 2;
rasbold@604	827	Node *argument2 = pop();
rasbold@604	828	Node *argument1 = pop();
rasbold@604	829
rasbold@604	830	Node* equals =
rasbold@604	831	_gvn.transform(new (C, 3) AryEqNode(control(),
rasbold@604	832	argument1,
rasbold@604	833	argument2)
rasbold@604	834	);
rasbold@604	835	push(equals);
rasbold@604	836	return true;
rasbold@604	837	}
rasbold@604	838
duke@435	839	// Java version of String.indexOf(constant string)
duke@435	840	// class StringDecl {
duke@435	841	// StringDecl(char[] ca) {
duke@435	842	// offset = 0;
duke@435	843	// count = ca.length;
duke@435	844	// value = ca;
duke@435	845	// }
duke@435	846	// int offset;
duke@435	847	// int count;
duke@435	848	// char[] value;
duke@435	849	// }
duke@435	850	//
duke@435	851	// static int string_indexOf_J(StringDecl string_object, char[] target_object,
duke@435	852	// int targetOffset, int cache_i, int md2) {
duke@435	853	// int cache = cache_i;
duke@435	854	// int sourceOffset = string_object.offset;
duke@435	855	// int sourceCount = string_object.count;
duke@435	856	// int targetCount = target_object.length;
duke@435	857	//
duke@435	858	// int targetCountLess1 = targetCount - 1;
duke@435	859	// int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
duke@435	860	//
duke@435	861	// char[] source = string_object.value;
duke@435	862	// char[] target = target_object;
duke@435	863	// int lastChar = target[targetCountLess1];
duke@435	864	//
duke@435	865	// outer_loop:
duke@435	866	// for (int i = sourceOffset; i < sourceEnd; ) {
duke@435	867	// int src = source[i + targetCountLess1];
duke@435	868	// if (src == lastChar) {
duke@435	869	// // With random strings and a 4-character alphabet,
duke@435	870	// // reverse matching at this point sets up 0.8% fewer
duke@435	871	// // frames, but (paradoxically) makes 0.3% more probes.
duke@435	872	// // Since those probes are nearer the lastChar probe,
duke@435	873	// // there is may be a net D$ win with reverse matching.
duke@435	874	// // But, reversing loop inhibits unroll of inner loop
duke@435	875	// // for unknown reason. So, does running outer loop from
duke@435	876	// // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
duke@435	877	// for (int j = 0; j < targetCountLess1; j++) {
duke@435	878	// if (target[targetOffset + j] != source[i+j]) {
duke@435	879	// if ((cache & (1 << source[i+j])) == 0) {
duke@435	880	// if (md2 < j+1) {
duke@435	881	// i += j+1;
duke@435	882	// continue outer_loop;
duke@435	883	// }
duke@435	884	// }
duke@435	885	// i += md2;
duke@435	886	// continue outer_loop;
duke@435	887	// }
duke@435	888	// }
duke@435	889	// return i - sourceOffset;
duke@435	890	// }
duke@435	891	// if ((cache & (1 << src)) == 0) {
duke@435	892	// i += targetCountLess1;
duke@435	893	// } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
duke@435	894	// i++;
duke@435	895	// }
duke@435	896	// return -1;
duke@435	897	// }
duke@435	898
duke@435	899	//------------------------------string_indexOf------------------------
duke@435	900	Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
duke@435	901	jint cache_i, jint md2_i) {
duke@435	902
duke@435	903	Node* no_ctrl = NULL;
duke@435	904	float likely = PROB_LIKELY(0.9);
duke@435	905	float unlikely = PROB_UNLIKELY(0.9);
duke@435	906
duke@435	907	const int value_offset = java_lang_String::value_offset_in_bytes();
duke@435	908	const int count_offset = java_lang_String::count_offset_in_bytes();
duke@435	909	const int offset_offset = java_lang_String::offset_offset_in_bytes();
duke@435	910
duke@435	911	ciInstanceKlass* klass = env()->String_klass();
duke@435	912	const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
duke@435	913	const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
duke@435	914
duke@435	915	Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
duke@435	916	Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
duke@435	917	Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
duke@435	918	Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
duke@435	919	Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
duke@435	920	Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
duke@435	921
kvn@599	922	Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
duke@435	923	jint target_length = target_array->length();
duke@435	924	const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
duke@435	925	const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
duke@435	926
duke@435	927	IdealKit kit(gvn(), control(), merged_memory());
duke@435	928	#define __ kit.
duke@435	929	Node* zero = __ ConI(0);
duke@435	930	Node* one = __ ConI(1);
duke@435	931	Node* cache = __ ConI(cache_i);
duke@435	932	Node* md2 = __ ConI(md2_i);
duke@435	933	Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
duke@435	934	Node* targetCount = __ ConI(target_length);
duke@435	935	Node* targetCountLess1 = __ ConI(target_length - 1);
duke@435	936	Node* targetOffset = __ ConI(targetOffset_i);
duke@435	937	Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
duke@435	938
duke@435	939	IdealVariable rtn(kit), i(kit), j(kit); __ declares_done();
duke@435	940	Node* outer_loop = __ make_label(2 /* goto */);
duke@435	941	Node* return_ = __ make_label(1);
duke@435	942
duke@435	943	__ set(rtn,__ ConI(-1));
duke@435	944	__ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
duke@435	945	Node* i2 = __ AddI(__ value(i), targetCountLess1);
duke@435	946	// pin to prohibit loading of "next iteration" value which may SEGV (rare)
duke@435	947	Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
duke@435	948	__ if_then(src, BoolTest::eq, lastChar, unlikely); {
duke@435	949	__ loop(j, zero, BoolTest::lt, targetCountLess1); {
duke@435	950	Node* tpj = __ AddI(targetOffset, __ value(j));
duke@435	951	Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
duke@435	952	Node* ipj = __ AddI(__ value(i), __ value(j));
duke@435	953	Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
duke@435	954	__ if_then(targ, BoolTest::ne, src2); {
duke@435	955	__ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
duke@435	956	__ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
duke@435	957	__ increment(i, __ AddI(__ value(j), one));
duke@435	958	__ goto_(outer_loop);
duke@435	959	} __ end_if(); __ dead(j);
duke@435	960	}__ end_if(); __ dead(j);
duke@435	961	__ increment(i, md2);
duke@435	962	__ goto_(outer_loop);
duke@435	963	}__ end_if();
duke@435	964	__ increment(j, one);
duke@435	965	}__ end_loop(); __ dead(j);
duke@435	966	__ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
duke@435	967	__ goto_(return_);
duke@435	968	}__ end_if();
duke@435	969	__ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
duke@435	970	__ increment(i, targetCountLess1);
duke@435	971	}__ end_if();
duke@435	972	__ increment(i, one);
duke@435	973	__ bind(outer_loop);
duke@435	974	}__ end_loop(); __ dead(i);
duke@435	975	__ bind(return_);
duke@435	976	__ drain_delay_transform();
duke@435	977
duke@435	978	set_control(__ ctrl());
duke@435	979	Node* result = __ value(rtn);
duke@435	980	#undef __
duke@435	981	C->set_has_loops(true);
duke@435	982	return result;
duke@435	983	}
duke@435	984
duke@435	985
duke@435	986	//------------------------------inline_string_indexOf------------------------
duke@435	987	bool LibraryCallKit::inline_string_indexOf() {
duke@435	988
duke@435	989	_sp += 2;
duke@435	990	Node *argument = pop(); // pop non-receiver first: it was pushed second
duke@435	991	Node *receiver = pop();
duke@435	992
duke@435	993	// don't intrinsify is argument isn't a constant string.
duke@435	994	if (!argument->is_Con()) {
duke@435	995	return false;
duke@435	996	}
duke@435	997	const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
duke@435	998	if (str_type == NULL) {
duke@435	999	return false;
duke@435	1000	}
duke@435	1001	ciInstanceKlass* klass = env()->String_klass();
duke@435	1002	ciObject* str_const = str_type->const_oop();
duke@435	1003	if (str_const == NULL || str_const->klass() != klass) {
duke@435	1004	return false;
duke@435	1005	}
duke@435	1006	ciInstance* str = str_const->as_instance();
duke@435	1007	assert(str != NULL, "must be instance");
duke@435	1008
duke@435	1009	const int value_offset = java_lang_String::value_offset_in_bytes();
duke@435	1010	const int count_offset = java_lang_String::count_offset_in_bytes();
duke@435	1011	const int offset_offset = java_lang_String::offset_offset_in_bytes();
duke@435	1012
duke@435	1013	ciObject* v = str->field_value_by_offset(value_offset).as_object();
duke@435	1014	int o = str->field_value_by_offset(offset_offset).as_int();
duke@435	1015	int c = str->field_value_by_offset(count_offset).as_int();
duke@435	1016	ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
duke@435	1017
duke@435	1018	// constant strings have no offset and count == length which
duke@435	1019	// simplifies the resulting code somewhat so lets optimize for that.
duke@435	1020	if (o != 0 || c != pat->length()) {
duke@435	1021	return false;
duke@435	1022	}
duke@435	1023
duke@435	1024	// Null check on self without removing any arguments. The argument
duke@435	1025	// null check technically happens in the wrong place, which can lead to
duke@435	1026	// invalid stack traces when string compare is inlined into a method
duke@435	1027	// which handles NullPointerExceptions.
duke@435	1028	_sp += 2;
duke@435	1029	receiver = do_null_check(receiver, T_OBJECT);
duke@435	1030	// No null check on the argument is needed since it's a constant String oop.
duke@435	1031	_sp -= 2;
duke@435	1032	if (stopped()) {
duke@435	1033	return true;
duke@435	1034	}
duke@435	1035
duke@435	1036	// The null string as a pattern always returns 0 (match at beginning of string)
duke@435	1037	if (c == 0) {
duke@435	1038	push(intcon(0));
duke@435	1039	return true;
duke@435	1040	}
duke@435	1041
duke@435	1042	jchar lastChar = pat->char_at(o + (c - 1));
duke@435	1043	int cache = 0;
duke@435	1044	int i;
duke@435	1045	for (i = 0; i < c - 1; i++) {
duke@435	1046	assert(i < pat->length(), "out of range");
duke@435	1047	cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
duke@435	1048	}
duke@435	1049
duke@435	1050	int md2 = c;
duke@435	1051	for (i = 0; i < c - 1; i++) {
duke@435	1052	assert(i < pat->length(), "out of range");
duke@435	1053	if (pat->char_at(o + i) == lastChar) {
duke@435	1054	md2 = (c - 1) - i;
duke@435	1055	}
duke@435	1056	}
duke@435	1057
duke@435	1058	Node* result = string_indexOf(receiver, pat, o, cache, md2);
duke@435	1059	push(result);
duke@435	1060	return true;
duke@435	1061	}
duke@435	1062
duke@435	1063	//--------------------------pop_math_arg--------------------------------
duke@435	1064	// Pop a double argument to a math function from the stack
duke@435	1065	// rounding it if necessary.
duke@435	1066	Node * LibraryCallKit::pop_math_arg() {
duke@435	1067	Node *arg = pop_pair();
duke@435	1068	if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
duke@435	1069	arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
duke@435	1070	return arg;
duke@435	1071	}
duke@435	1072
duke@435	1073	//------------------------------inline_trig----------------------------------
duke@435	1074	// Inline sin/cos/tan instructions, if possible. If rounding is required, do
duke@435	1075	// argument reduction which will turn into a fast/slow diamond.
duke@435	1076	bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
duke@435	1077	_sp += arg_size(); // restore stack pointer
duke@435	1078	Node* arg = pop_math_arg();
duke@435	1079	Node* trig = NULL;
duke@435	1080
duke@435	1081	switch (id) {
duke@435	1082	case vmIntrinsics::_dsin:
duke@435	1083	trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
duke@435	1084	break;
duke@435	1085	case vmIntrinsics::_dcos:
duke@435	1086	trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
duke@435	1087	break;
duke@435	1088	case vmIntrinsics::_dtan:
duke@435	1089	trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
duke@435	1090	break;
duke@435	1091	default:
duke@435	1092	assert(false, "bad intrinsic was passed in");
duke@435	1093	return false;
duke@435	1094	}
duke@435	1095
duke@435	1096	// Rounding required? Check for argument reduction!
duke@435	1097	if( Matcher::strict_fp_requires_explicit_rounding ) {
duke@435	1098
duke@435	1099	static const double pi_4 = 0.7853981633974483;
duke@435	1100	static const double neg_pi_4 = -0.7853981633974483;
duke@435	1101	// pi/2 in 80-bit extended precision
duke@435	1102	// 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	1103	// -pi/2 in 80-bit extended precision
duke@435	1104	// 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	1105	// Cutoff value for using this argument reduction technique
duke@435	1106	//static const double pi_2_minus_epsilon = 1.564660403643354;
duke@435	1107	//static const double neg_pi_2_plus_epsilon = -1.564660403643354;
duke@435	1108
duke@435	1109	// Pseudocode for sin:
duke@435	1110	// if (x <= Math.PI / 4.0) {
duke@435	1111	// if (x >= -Math.PI / 4.0) return fsin(x);
duke@435	1112	// if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
duke@435	1113	// } else {
duke@435	1114	// if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
duke@435	1115	// }
duke@435	1116	// return StrictMath.sin(x);
duke@435	1117
duke@435	1118	// Pseudocode for cos:
duke@435	1119	// if (x <= Math.PI / 4.0) {
duke@435	1120	// if (x >= -Math.PI / 4.0) return fcos(x);
duke@435	1121	// if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
duke@435	1122	// } else {
duke@435	1123	// if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
duke@435	1124	// }
duke@435	1125	// return StrictMath.cos(x);
duke@435	1126
duke@435	1127	// Actually, sticking in an 80-bit Intel value into C2 will be tough; it
duke@435	1128	// requires a special machine instruction to load it. Instead we'll try
duke@435	1129	// the 'easy' case. If we really need the extra range +/- PI/2 we'll
duke@435	1130	// probably do the math inside the SIN encoding.
duke@435	1131
duke@435	1132	// Make the merge point
duke@435	1133	RegionNode *r = new (C, 3) RegionNode(3);
duke@435	1134	Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
duke@435	1135
duke@435	1136	// Flatten arg so we need only 1 test
duke@435	1137	Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
duke@435	1138	// Node for PI/4 constant
duke@435	1139	Node *pi4 = makecon(TypeD::make(pi_4));
duke@435	1140	// Check PI/4 : abs(arg)
duke@435	1141	Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
duke@435	1142	// Check: If PI/4 < abs(arg) then go slow
duke@435	1143	Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
duke@435	1144	// Branch either way
duke@435	1145	IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
duke@435	1146	set_control(opt_iff(r,iff));
duke@435	1147
duke@435	1148	// Set fast path result
duke@435	1149	phi->init_req(2,trig);
duke@435	1150
duke@435	1151	// Slow path - non-blocking leaf call
duke@435	1152	Node* call = NULL;
duke@435	1153	switch (id) {
duke@435	1154	case vmIntrinsics::_dsin:
duke@435	1155	call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
duke@435	1156	CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
duke@435	1157	"Sin", NULL, arg, top());
duke@435	1158	break;
duke@435	1159	case vmIntrinsics::_dcos:
duke@435	1160	call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
duke@435	1161	CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
duke@435	1162	"Cos", NULL, arg, top());
duke@435	1163	break;
duke@435	1164	case vmIntrinsics::_dtan:
duke@435	1165	call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
duke@435	1166	CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
duke@435	1167	"Tan", NULL, arg, top());
duke@435	1168	break;
duke@435	1169	}
duke@435	1170	assert(control()->in(0) == call, "");
duke@435	1171	Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
duke@435	1172	r->init_req(1,control());
duke@435	1173	phi->init_req(1,slow_result);
duke@435	1174
duke@435	1175	// Post-merge
duke@435	1176	set_control(_gvn.transform(r));
duke@435	1177	record_for_igvn(r);
duke@435	1178	trig = _gvn.transform(phi);
duke@435	1179
duke@435	1180	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	1181	}
duke@435	1182	// Push result back on JVM stack
duke@435	1183	push_pair(trig);
duke@435	1184	return true;
duke@435	1185	}
duke@435	1186
duke@435	1187	//------------------------------inline_sqrt-------------------------------------
duke@435	1188	// Inline square root instruction, if possible.
duke@435	1189	bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
duke@435	1190	assert(id == vmIntrinsics::_dsqrt, "Not square root");
duke@435	1191	_sp += arg_size(); // restore stack pointer
duke@435	1192	push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
duke@435	1193	return true;
duke@435	1194	}
duke@435	1195
duke@435	1196	//------------------------------inline_abs-------------------------------------
duke@435	1197	// Inline absolute value instruction, if possible.
duke@435	1198	bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
duke@435	1199	assert(id == vmIntrinsics::_dabs, "Not absolute value");
duke@435	1200	_sp += arg_size(); // restore stack pointer
duke@435	1201	push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
duke@435	1202	return true;
duke@435	1203	}
duke@435	1204
duke@435	1205	//------------------------------inline_exp-------------------------------------
duke@435	1206	// Inline exp instructions, if possible. The Intel hardware only misses
duke@435	1207	// really odd corner cases (+/- Infinity). Just uncommon-trap them.
duke@435	1208	bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
duke@435	1209	assert(id == vmIntrinsics::_dexp, "Not exp");
duke@435	1210
duke@435	1211	// If this inlining ever returned NaN in the past, we do not intrinsify it
duke@435	1212	// every again. NaN results requires StrictMath.exp handling.
duke@435	1213	if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@435	1214
duke@435	1215	// Do not intrinsify on older platforms which lack cmove.
duke@435	1216	if (ConditionalMoveLimit == 0) return false;
duke@435	1217
duke@435	1218	_sp += arg_size(); // restore stack pointer
duke@435	1219	Node *x = pop_math_arg();
duke@435	1220	Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
duke@435	1221
duke@435	1222	//-------------------
duke@435	1223	//result=(result.isNaN())? StrictMath::exp():result;
duke@435	1224	// Check: If isNaN() by checking result!=result? then go to Strict Math
duke@435	1225	Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
duke@435	1226	// Build the boolean node
duke@435	1227	Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
duke@435	1228
duke@435	1229	{ BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
duke@435	1230	// End the current control-flow path
duke@435	1231	push_pair(x);
duke@435	1232	// Math.exp intrinsic returned a NaN, which requires StrictMath.exp
duke@435	1233	// to handle. Recompile without intrinsifying Math.exp
duke@435	1234	uncommon_trap(Deoptimization::Reason_intrinsic,
duke@435	1235	Deoptimization::Action_make_not_entrant);
duke@435	1236	}
duke@435	1237
duke@435	1238	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	1239
duke@435	1240	push_pair(result);
duke@435	1241
duke@435	1242	return true;
duke@435	1243	}
duke@435	1244
duke@435	1245	//------------------------------inline_pow-------------------------------------
duke@435	1246	// Inline power instructions, if possible.
duke@435	1247	bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
duke@435	1248	assert(id == vmIntrinsics::_dpow, "Not pow");
duke@435	1249
duke@435	1250	// If this inlining ever returned NaN in the past, we do not intrinsify it
duke@435	1251	// every again. NaN results requires StrictMath.pow handling.
duke@435	1252	if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@435	1253
duke@435	1254	// Do not intrinsify on older platforms which lack cmove.
duke@435	1255	if (ConditionalMoveLimit == 0) return false;
duke@435	1256
duke@435	1257	// Pseudocode for pow
duke@435	1258	// if (x <= 0.0) {
duke@435	1259	// if ((double)((int)y)==y) { // if y is int
duke@435	1260	// result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
duke@435	1261	// } else {
duke@435	1262	// result = NaN;
duke@435	1263	// }
duke@435	1264	// } else {
duke@435	1265	// result = DPow(x,y);
duke@435	1266	// }
duke@435	1267	// if (result != result)? {
duke@435	1268	// ucommon_trap();
duke@435	1269	// }
duke@435	1270	// return result;
duke@435	1271
duke@435	1272	_sp += arg_size(); // restore stack pointer
duke@435	1273	Node* y = pop_math_arg();
duke@435	1274	Node* x = pop_math_arg();
duke@435	1275
duke@435	1276	Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
duke@435	1277
duke@435	1278	// Short form: if not top-level (i.e., Math.pow but inlining Math.pow
duke@435	1279	// inside of something) then skip the fancy tests and just check for
duke@435	1280	// NaN result.
duke@435	1281	Node *result = NULL;
duke@435	1282	if( jvms()->depth() >= 1 ) {
duke@435	1283	result = fast_result;
duke@435	1284	} else {
duke@435	1285
duke@435	1286	// Set the merge point for If node with condition of (x <= 0.0)
duke@435	1287	// There are four possible paths to region node and phi node
duke@435	1288	RegionNode *r = new (C, 4) RegionNode(4);
duke@435	1289	Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
duke@435	1290
duke@435	1291	// Build the first if node: if (x <= 0.0)
duke@435	1292	// Node for 0 constant
duke@435	1293	Node *zeronode = makecon(TypeD::ZERO);
duke@435	1294	// Check x:0
duke@435	1295	Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
duke@435	1296	// Check: If (x<=0) then go complex path
duke@435	1297	Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
duke@435	1298	// Branch either way
duke@435	1299	IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
duke@435	1300	Node *opt_test = _gvn.transform(if1);
duke@435	1301	//assert( opt_test->is_If(), "Expect an IfNode");
duke@435	1302	IfNode *opt_if1 = (IfNode*)opt_test;
duke@435	1303	// Fast path taken; set region slot 3
duke@435	1304	Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
duke@435	1305	r->init_req(3,fast_taken); // Capture fast-control
duke@435	1306
duke@435	1307	// Fast path not-taken, i.e. slow path
duke@435	1308	Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
duke@435	1309
duke@435	1310	// Set fast path result
duke@435	1311	Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
duke@435	1312	phi->init_req(3, fast_result);
duke@435	1313
duke@435	1314	// Complex path
duke@435	1315	// Build the second if node (if y is int)
duke@435	1316	// Node for (int)y
duke@435	1317	Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
duke@435	1318	// Node for (double)((int) y)
duke@435	1319	Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
duke@435	1320	// Check (double)((int) y) : y
duke@435	1321	Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
duke@435	1322	// Check if (y isn't int) then go to slow path
duke@435	1323
duke@435	1324	Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
duke@435	1325	// Branch eith way
duke@435	1326	IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
duke@435	1327	Node *slow_path = opt_iff(r,if2); // Set region path 2
duke@435	1328
duke@435	1329	// Calculate DPow(abs(x), y)*(1 & (int)y)
duke@435	1330	// Node for constant 1
duke@435	1331	Node *conone = intcon(1);
duke@435	1332	// 1& (int)y
duke@435	1333	Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
duke@435	1334	// zero node
duke@435	1335	Node *conzero = intcon(0);
duke@435	1336	// Check (1&(int)y)==0?
duke@435	1337	Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
duke@435	1338	// Check if (1&(int)y)!=0?, if so the result is negative
duke@435	1339	Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
duke@435	1340	// abs(x)
duke@435	1341	Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
duke@435	1342	// abs(x)^y
duke@435	1343	Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
duke@435	1344	// -abs(x)^y
duke@435	1345	Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
duke@435	1346	// (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
duke@435	1347	Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
duke@435	1348	// Set complex path fast result
duke@435	1349	phi->init_req(2, signresult);
duke@435	1350
duke@435	1351	static const jlong nan_bits = CONST64(0x7ff8000000000000);
duke@435	1352	Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
duke@435	1353	r->init_req(1,slow_path);
duke@435	1354	phi->init_req(1,slow_result);
duke@435	1355
duke@435	1356	// Post merge
duke@435	1357	set_control(_gvn.transform(r));
duke@435	1358	record_for_igvn(r);
duke@435	1359	result=_gvn.transform(phi);
duke@435	1360	}
duke@435	1361
duke@435	1362	//-------------------
duke@435	1363	//result=(result.isNaN())? uncommon_trap():result;
duke@435	1364	// Check: If isNaN() by checking result!=result? then go to Strict Math
duke@435	1365	Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
duke@435	1366	// Build the boolean node
duke@435	1367	Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
duke@435	1368
duke@435	1369	{ BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
duke@435	1370	// End the current control-flow path
duke@435	1371	push_pair(x);
duke@435	1372	push_pair(y);
duke@435	1373	// Math.pow intrinsic returned a NaN, which requires StrictMath.pow
duke@435	1374	// to handle. Recompile without intrinsifying Math.pow.
duke@435	1375	uncommon_trap(Deoptimization::Reason_intrinsic,
duke@435	1376	Deoptimization::Action_make_not_entrant);
duke@435	1377	}
duke@435	1378
duke@435	1379	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	1380
duke@435	1381	push_pair(result);
duke@435	1382
duke@435	1383	return true;
duke@435	1384	}
duke@435	1385
duke@435	1386	//------------------------------inline_trans-------------------------------------
duke@435	1387	// Inline transcendental instructions, if possible. The Intel hardware gets
duke@435	1388	// these right, no funny corner cases missed.
duke@435	1389	bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
duke@435	1390	_sp += arg_size(); // restore stack pointer
duke@435	1391	Node* arg = pop_math_arg();
duke@435	1392	Node* trans = NULL;
duke@435	1393
duke@435	1394	switch (id) {
duke@435	1395	case vmIntrinsics::_dlog:
duke@435	1396	trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
duke@435	1397	break;
duke@435	1398	case vmIntrinsics::_dlog10:
duke@435	1399	trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
duke@435	1400	break;
duke@435	1401	default:
duke@435	1402	assert(false, "bad intrinsic was passed in");
duke@435	1403	return false;
duke@435	1404	}
duke@435	1405
duke@435	1406	// Push result back on JVM stack
duke@435	1407	push_pair(trans);
duke@435	1408	return true;
duke@435	1409	}
duke@435	1410
duke@435	1411	//------------------------------runtime_math-----------------------------
duke@435	1412	bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
duke@435	1413	Node* a = NULL;
duke@435	1414	Node* b = NULL;
duke@435	1415
duke@435	1416	assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
duke@435	1417	"must be (DD)D or (D)D type");
duke@435	1418
duke@435	1419	// Inputs
duke@435	1420	_sp += arg_size(); // restore stack pointer
duke@435	1421	if (call_type == OptoRuntime::Math_DD_D_Type()) {
duke@435	1422	b = pop_math_arg();
duke@435	1423	}
duke@435	1424	a = pop_math_arg();
duke@435	1425
duke@435	1426	const TypePtr* no_memory_effects = NULL;
duke@435	1427	Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
duke@435	1428	no_memory_effects,
duke@435	1429	a, top(), b, b ? top() : NULL);
duke@435	1430	Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
duke@435	1431	#ifdef ASSERT
duke@435	1432	Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
duke@435	1433	assert(value_top == top(), "second value must be top");
duke@435	1434	#endif
duke@435	1435
duke@435	1436	push_pair(value);
duke@435	1437	return true;
duke@435	1438	}
duke@435	1439
duke@435	1440	//------------------------------inline_math_native-----------------------------
duke@435	1441	bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
duke@435	1442	switch (id) {
duke@435	1443	// These intrinsics are not properly supported on all hardware
duke@435	1444	case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
duke@435	1445	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
duke@435	1446	case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
duke@435	1447	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
duke@435	1448	case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
duke@435	1449	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
duke@435	1450
duke@435	1451	case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
duke@435	1452	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
duke@435	1453	case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
duke@435	1454	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
duke@435	1455
duke@435	1456	// These intrinsics are supported on all hardware
duke@435	1457	case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
duke@435	1458	case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
duke@435	1459
duke@435	1460	// These intrinsics don't work on X86. The ad implementation doesn't
duke@435	1461	// handle NaN's properly. Instead of returning infinity, the ad
duke@435	1462	// implementation returns a NaN on overflow. See bug: 6304089
duke@435	1463	// Once the ad implementations are fixed, change the code below
duke@435	1464	// to match the intrinsics above
duke@435	1465
duke@435	1466	case vmIntrinsics::_dexp: return
duke@435	1467	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
duke@435	1468	case vmIntrinsics::_dpow: return
duke@435	1469	runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
duke@435	1470
duke@435	1471	// These intrinsics are not yet correctly implemented
duke@435	1472	case vmIntrinsics::_datan2:
duke@435	1473	return false;
duke@435	1474
duke@435	1475	default:
duke@435	1476	ShouldNotReachHere();
duke@435	1477	return false;
duke@435	1478	}
duke@435	1479	}
duke@435	1480
duke@435	1481	static bool is_simple_name(Node* n) {
duke@435	1482	return (n->req() == 1 // constant
duke@435	1483	|| (n->is_Type() && n->as_Type()->type()->singleton())
duke@435	1484	|| n->is_Proj() // parameter or return value
duke@435	1485	|| n->is_Phi() // local of some sort
duke@435	1486	);
duke@435	1487	}
duke@435	1488
duke@435	1489	//----------------------------inline_min_max-----------------------------------
duke@435	1490	bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
duke@435	1491	push(generate_min_max(id, argument(0), argument(1)));
duke@435	1492
duke@435	1493	return true;
duke@435	1494	}
duke@435	1495
duke@435	1496	Node*
duke@435	1497	LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
duke@435	1498	// These are the candidate return value:
duke@435	1499	Node* xvalue = x0;
duke@435	1500	Node* yvalue = y0;
duke@435	1501
duke@435	1502	if (xvalue == yvalue) {
duke@435	1503	return xvalue;
duke@435	1504	}
duke@435	1505
duke@435	1506	bool want_max = (id == vmIntrinsics::_max);
duke@435	1507
duke@435	1508	const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
duke@435	1509	const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
duke@435	1510	if (txvalue == NULL || tyvalue == NULL) return top();
duke@435	1511	// This is not really necessary, but it is consistent with a
duke@435	1512	// hypothetical MaxINode::Value method:
duke@435	1513	int widen = MAX2(txvalue->_widen, tyvalue->_widen);
duke@435	1514
duke@435	1515	// %%% This folding logic should (ideally) be in a different place.
duke@435	1516	// Some should be inside IfNode, and there to be a more reliable
duke@435	1517	// transformation of ?: style patterns into cmoves. We also want
duke@435	1518	// more powerful optimizations around cmove and min/max.
duke@435	1519
duke@435	1520	// Try to find a dominating comparison of these guys.
duke@435	1521	// It can simplify the index computation for Arrays.copyOf
duke@435	1522	// and similar uses of System.arraycopy.
duke@435	1523	// First, compute the normalized version of CmpI(x, y).
duke@435	1524	int cmp_op = Op_CmpI;
duke@435	1525	Node* xkey = xvalue;
duke@435	1526	Node* ykey = yvalue;
duke@435	1527	Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
duke@435	1528	if (ideal_cmpxy->is_Cmp()) {
duke@435	1529	// E.g., if we have CmpI(length - offset, count),
duke@435	1530	// it might idealize to CmpI(length, count + offset)
duke@435	1531	cmp_op = ideal_cmpxy->Opcode();
duke@435	1532	xkey = ideal_cmpxy->in(1);
duke@435	1533	ykey = ideal_cmpxy->in(2);
duke@435	1534	}
duke@435	1535
duke@435	1536	// Start by locating any relevant comparisons.
duke@435	1537	Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
duke@435	1538	Node* cmpxy = NULL;
duke@435	1539	Node* cmpyx = NULL;
duke@435	1540	for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
duke@435	1541	Node* cmp = start_from->fast_out(k);
duke@435	1542	if (cmp->outcnt() > 0 && // must have prior uses
duke@435	1543	cmp->in(0) == NULL && // must be context-independent
duke@435	1544	cmp->Opcode() == cmp_op) { // right kind of compare
duke@435	1545	if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
duke@435	1546	if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
duke@435	1547	}
duke@435	1548	}
duke@435	1549
duke@435	1550	const int NCMPS = 2;
duke@435	1551	Node* cmps[NCMPS] = { cmpxy, cmpyx };
duke@435	1552	int cmpn;
duke@435	1553	for (cmpn = 0; cmpn < NCMPS; cmpn++) {
duke@435	1554	if (cmps[cmpn] != NULL) break; // find a result
duke@435	1555	}
duke@435	1556	if (cmpn < NCMPS) {
duke@435	1557	// Look for a dominating test that tells us the min and max.
duke@435	1558	int depth = 0; // Limit search depth for speed
duke@435	1559	Node* dom = control();
duke@435	1560	for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
duke@435	1561	if (++depth >= 100) break;
duke@435	1562	Node* ifproj = dom;
duke@435	1563	if (!ifproj->is_Proj()) continue;
duke@435	1564	Node* iff = ifproj->in(0);
duke@435	1565	if (!iff->is_If()) continue;
duke@435	1566	Node* bol = iff->in(1);
duke@435	1567	if (!bol->is_Bool()) continue;
duke@435	1568	Node* cmp = bol->in(1);
duke@435	1569	if (cmp == NULL) continue;
duke@435	1570	for (cmpn = 0; cmpn < NCMPS; cmpn++)
duke@435	1571	if (cmps[cmpn] == cmp) break;
duke@435	1572	if (cmpn == NCMPS) continue;
duke@435	1573	BoolTest::mask btest = bol->as_Bool()->_test._test;
duke@435	1574	if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
duke@435	1575	if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
duke@435	1576	// At this point, we know that 'x btest y' is true.
duke@435	1577	switch (btest) {
duke@435	1578	case BoolTest::eq:
duke@435	1579	// They are proven equal, so we can collapse the min/max.
duke@435	1580	// Either value is the answer. Choose the simpler.
duke@435	1581	if (is_simple_name(yvalue) && !is_simple_name(xvalue))
duke@435	1582	return yvalue;
duke@435	1583	return xvalue;
duke@435	1584	case BoolTest::lt: // x < y
duke@435	1585	case BoolTest::le: // x <= y
duke@435	1586	return (want_max ? yvalue : xvalue);
duke@435	1587	case BoolTest::gt: // x > y
duke@435	1588	case BoolTest::ge: // x >= y
duke@435	1589	return (want_max ? xvalue : yvalue);
duke@435	1590	}
duke@435	1591	}
duke@435	1592	}
duke@435	1593
duke@435	1594	// We failed to find a dominating test.
duke@435	1595	// Let's pick a test that might GVN with prior tests.
duke@435	1596	Node* best_bol = NULL;
duke@435	1597	BoolTest::mask best_btest = BoolTest::illegal;
duke@435	1598	for (cmpn = 0; cmpn < NCMPS; cmpn++) {
duke@435	1599	Node* cmp = cmps[cmpn];
duke@435	1600	if (cmp == NULL) continue;
duke@435	1601	for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
duke@435	1602	Node* bol = cmp->fast_out(j);
duke@435	1603	if (!bol->is_Bool()) continue;
duke@435	1604	BoolTest::mask btest = bol->as_Bool()->_test._test;
duke@435	1605	if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
duke@435	1606	if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
duke@435	1607	if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
duke@435	1608	best_bol = bol->as_Bool();
duke@435	1609	best_btest = btest;
duke@435	1610	}
duke@435	1611	}
duke@435	1612	}
duke@435	1613
duke@435	1614	Node* answer_if_true = NULL;
duke@435	1615	Node* answer_if_false = NULL;
duke@435	1616	switch (best_btest) {
duke@435	1617	default:
duke@435	1618	if (cmpxy == NULL)
duke@435	1619	cmpxy = ideal_cmpxy;
duke@435	1620	best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
duke@435	1621	// and fall through:
duke@435	1622	case BoolTest::lt: // x < y
duke@435	1623	case BoolTest::le: // x <= y
duke@435	1624	answer_if_true = (want_max ? yvalue : xvalue);
duke@435	1625	answer_if_false = (want_max ? xvalue : yvalue);
duke@435	1626	break;
duke@435	1627	case BoolTest::gt: // x > y
duke@435	1628	case BoolTest::ge: // x >= y
duke@435	1629	answer_if_true = (want_max ? xvalue : yvalue);
duke@435	1630	answer_if_false = (want_max ? yvalue : xvalue);
duke@435	1631	break;
duke@435	1632	}
duke@435	1633
duke@435	1634	jint hi, lo;
duke@435	1635	if (want_max) {
duke@435	1636	// We can sharpen the minimum.
duke@435	1637	hi = MAX2(txvalue->_hi, tyvalue->_hi);
duke@435	1638	lo = MAX2(txvalue->_lo, tyvalue->_lo);
duke@435	1639	} else {
duke@435	1640	// We can sharpen the maximum.
duke@435	1641	hi = MIN2(txvalue->_hi, tyvalue->_hi);
duke@435	1642	lo = MIN2(txvalue->_lo, tyvalue->_lo);
duke@435	1643	}
duke@435	1644
duke@435	1645	// Use a flow-free graph structure, to avoid creating excess control edges
duke@435	1646	// which could hinder other optimizations.
duke@435	1647	// Since Math.min/max is often used with arraycopy, we want
duke@435	1648	// tightly_coupled_allocation to be able to see beyond min/max expressions.
duke@435	1649	Node* cmov = CMoveNode::make(C, NULL, best_bol,
duke@435	1650	answer_if_false, answer_if_true,
duke@435	1651	TypeInt::make(lo, hi, widen));
duke@435	1652
duke@435	1653	return _gvn.transform(cmov);
duke@435	1654
duke@435	1655	/*
duke@435	1656	// This is not as desirable as it may seem, since Min and Max
duke@435	1657	// nodes do not have a full set of optimizations.
duke@435	1658	// And they would interfere, anyway, with 'if' optimizations
duke@435	1659	// and with CMoveI canonical forms.
duke@435	1660	switch (id) {
duke@435	1661	case vmIntrinsics::_min:
duke@435	1662	result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
duke@435	1663	case vmIntrinsics::_max:
duke@435	1664	result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
duke@435	1665	default:
duke@435	1666	ShouldNotReachHere();
duke@435	1667	}
duke@435	1668	*/
duke@435	1669	}
duke@435	1670
duke@435	1671	inline int
duke@435	1672	LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
duke@435	1673	const TypePtr* base_type = TypePtr::NULL_PTR;
duke@435	1674	if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
duke@435	1675	if (base_type == NULL) {
duke@435	1676	// Unknown type.
duke@435	1677	return Type::AnyPtr;
duke@435	1678	} else if (base_type == TypePtr::NULL_PTR) {
duke@435	1679	// Since this is a NULL+long form, we have to switch to a rawptr.
duke@435	1680	base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
duke@435	1681	offset = MakeConX(0);
duke@435	1682	return Type::RawPtr;
duke@435	1683	} else if (base_type->base() == Type::RawPtr) {
duke@435	1684	return Type::RawPtr;
duke@435	1685	} else if (base_type->isa_oopptr()) {
duke@435	1686	// Base is never null => always a heap address.
duke@435	1687	if (base_type->ptr() == TypePtr::NotNull) {
duke@435	1688	return Type::OopPtr;
duke@435	1689	}
duke@435	1690	// Offset is small => always a heap address.
duke@435	1691	const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
duke@435	1692	if (offset_type != NULL &&
duke@435	1693	base_type->offset() == 0 && // (should always be?)
duke@435	1694	offset_type->_lo >= 0 &&
duke@435	1695	!MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
duke@435	1696	return Type::OopPtr;
duke@435	1697	}
duke@435	1698	// Otherwise, it might either be oop+off or NULL+addr.
duke@435	1699	return Type::AnyPtr;
duke@435	1700	} else {
duke@435	1701	// No information:
duke@435	1702	return Type::AnyPtr;
duke@435	1703	}
duke@435	1704	}
duke@435	1705
duke@435	1706	inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
duke@435	1707	int kind = classify_unsafe_addr(base, offset);
duke@435	1708	if (kind == Type::RawPtr) {
duke@435	1709	return basic_plus_adr(top(), base, offset);
duke@435	1710	} else {
duke@435	1711	return basic_plus_adr(base, offset);
duke@435	1712	}
duke@435	1713	}
duke@435	1714
duke@435	1715	//----------------------------inline_reverseBytes_int/long-------------------
duke@435	1716	// inline Int.reverseBytes(int)
duke@435	1717	// inline Long.reverseByes(long)
duke@435	1718	bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
duke@435	1719	assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes");
duke@435	1720	if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
duke@435	1721	if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
duke@435	1722	_sp += arg_size(); // restore stack pointer
duke@435	1723	switch (id) {
duke@435	1724	case vmIntrinsics::_reverseBytes_i:
duke@435	1725	push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
duke@435	1726	break;
duke@435	1727	case vmIntrinsics::_reverseBytes_l:
duke@435	1728	push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
duke@435	1729	break;
duke@435	1730	default:
duke@435	1731	;
duke@435	1732	}
duke@435	1733	return true;
duke@435	1734	}
duke@435	1735
duke@435	1736	//----------------------------inline_unsafe_access----------------------------
duke@435	1737
duke@435	1738	const static BasicType T_ADDRESS_HOLDER = T_LONG;
duke@435	1739
duke@435	1740	// Interpret Unsafe.fieldOffset cookies correctly:
duke@435	1741	extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
duke@435	1742
duke@435	1743	bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
duke@435	1744	if (callee()->is_static()) return false; // caller must have the capability!
duke@435	1745
duke@435	1746	#ifndef PRODUCT
duke@435	1747	{
duke@435	1748	ResourceMark rm;
duke@435	1749	// Check the signatures.
duke@435	1750	ciSignature* sig = signature();
duke@435	1751	#ifdef ASSERT
duke@435	1752	if (!is_store) {
duke@435	1753	// Object getObject(Object base, int/long offset), etc.
duke@435	1754	BasicType rtype = sig->return_type()->basic_type();
duke@435	1755	if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
duke@435	1756	rtype = T_ADDRESS; // it is really a C void*
duke@435	1757	assert(rtype == type, "getter must return the expected value");
duke@435	1758	if (!is_native_ptr) {
duke@435	1759	assert(sig->count() == 2, "oop getter has 2 arguments");
duke@435	1760	assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
duke@435	1761	assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
duke@435	1762	} else {
duke@435	1763	assert(sig->count() == 1, "native getter has 1 argument");
duke@435	1764	assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
duke@435	1765	}
duke@435	1766	} else {
duke@435	1767	// void putObject(Object base, int/long offset, Object x), etc.
duke@435	1768	assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
duke@435	1769	if (!is_native_ptr) {
duke@435	1770	assert(sig->count() == 3, "oop putter has 3 arguments");
duke@435	1771	assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
duke@435	1772	assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
duke@435	1773	} else {
duke@435	1774	assert(sig->count() == 2, "native putter has 2 arguments");
duke@435	1775	assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
duke@435	1776	}
duke@435	1777	BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
duke@435	1778	if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
duke@435	1779	vtype = T_ADDRESS; // it is really a C void*
duke@435	1780	assert(vtype == type, "putter must accept the expected value");
duke@435	1781	}
duke@435	1782	#endif // ASSERT
duke@435	1783	}
duke@435	1784	#endif //PRODUCT
duke@435	1785
duke@435	1786	C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@435	1787
duke@435	1788	int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
duke@435	1789
duke@435	1790	// Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
duke@435	1791	int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
duke@435	1792
duke@435	1793	debug_only(int saved_sp = _sp);
duke@435	1794	_sp += nargs;
duke@435	1795
duke@435	1796	Node* val;
duke@435	1797	debug_only(val = (Node*)(uintptr_t)-1);
duke@435	1798
duke@435	1799
duke@435	1800	if (is_store) {
duke@435	1801	// Get the value being stored. (Pop it first; it was pushed last.)
duke@435	1802	switch (type) {
duke@435	1803	case T_DOUBLE:
duke@435	1804	case T_LONG:
duke@435	1805	case T_ADDRESS:
duke@435	1806	val = pop_pair();
duke@435	1807	break;
duke@435	1808	default:
duke@435	1809	val = pop();
duke@435	1810	}
duke@435	1811	}
duke@435	1812
duke@435	1813	// Build address expression. See the code in inline_unsafe_prefetch.
duke@435	1814	Node *adr;
duke@435	1815	Node *heap_base_oop = top();
duke@435	1816	if (!is_native_ptr) {
duke@435	1817	// The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
duke@435	1818	Node* offset = pop_pair();
duke@435	1819	// The base is either a Java object or a value produced by Unsafe.staticFieldBase
duke@435	1820	Node* base = pop();
duke@435	1821	// We currently rely on the cookies produced by Unsafe.xxxFieldOffset
duke@435	1822	// to be plain byte offsets, which are also the same as those accepted
duke@435	1823	// by oopDesc::field_base.
duke@435	1824	assert(Unsafe_field_offset_to_byte_offset(11) == 11,
duke@435	1825	"fieldOffset must be byte-scaled");
duke@435	1826	// 32-bit machines ignore the high half!
duke@435	1827	offset = ConvL2X(offset);
duke@435	1828	adr = make_unsafe_address(base, offset);
duke@435	1829	heap_base_oop = base;
duke@435	1830	} else {
duke@435	1831	Node* ptr = pop_pair();
duke@435	1832	// Adjust Java long to machine word:
duke@435	1833	ptr = ConvL2X(ptr);
duke@435	1834	adr = make_unsafe_address(NULL, ptr);
duke@435	1835	}
duke@435	1836
duke@435	1837	// Pop receiver last: it was pushed first.
duke@435	1838	Node *receiver = pop();
duke@435	1839
duke@435	1840	assert(saved_sp == _sp, "must have correct argument count");
duke@435	1841
duke@435	1842	const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
duke@435	1843
duke@435	1844	// First guess at the value type.
duke@435	1845	const Type *value_type = Type::get_const_basic_type(type);
duke@435	1846
duke@435	1847	// Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
duke@435	1848	// there was not enough information to nail it down.
duke@435	1849	Compile::AliasType* alias_type = C->alias_type(adr_type);
duke@435	1850	assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
duke@435	1851
duke@435	1852	// We will need memory barriers unless we can determine a unique
duke@435	1853	// alias category for this reference. (Note: If for some reason
duke@435	1854	// the barriers get omitted and the unsafe reference begins to "pollute"
duke@435	1855	// the alias analysis of the rest of the graph, either Compile::can_alias
duke@435	1856	// or Compile::must_alias will throw a diagnostic assert.)
duke@435	1857	bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
duke@435	1858
duke@435	1859	if (!is_store && type == T_OBJECT) {
duke@435	1860	// Attempt to infer a sharper value type from the offset and base type.
duke@435	1861	ciKlass* sharpened_klass = NULL;
duke@435	1862
duke@435	1863	// See if it is an instance field, with an object type.
duke@435	1864	if (alias_type->field() != NULL) {
duke@435	1865	assert(!is_native_ptr, "native pointer op cannot use a java address");
duke@435	1866	if (alias_type->field()->type()->is_klass()) {
duke@435	1867	sharpened_klass = alias_type->field()->type()->as_klass();
duke@435	1868	}
duke@435	1869	}
duke@435	1870
duke@435	1871	// See if it is a narrow oop array.
duke@435	1872	if (adr_type->isa_aryptr()) {
coleenp@548	1873	if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes(type)) {
duke@435	1874	const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
duke@435	1875	if (elem_type != NULL) {
duke@435	1876	sharpened_klass = elem_type->klass();
duke@435	1877	}
duke@435	1878	}
duke@435	1879	}
duke@435	1880
duke@435	1881	if (sharpened_klass != NULL) {
duke@435	1882	const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
duke@435	1883
duke@435	1884	// Sharpen the value type.
duke@435	1885	value_type = tjp;
duke@435	1886
duke@435	1887	#ifndef PRODUCT
duke@435	1888	if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
duke@435	1889	tty->print(" from base type: "); adr_type->dump();
duke@435	1890	tty->print(" sharpened value: "); value_type->dump();
duke@435	1891	}
duke@435	1892	#endif
duke@435	1893	}
duke@435	1894	}
duke@435	1895
duke@435	1896	// Null check on self without removing any arguments. The argument
duke@435	1897	// null check technically happens in the wrong place, which can lead to
duke@435	1898	// invalid stack traces when the primitive is inlined into a method
duke@435	1899	// which handles NullPointerExceptions.
duke@435	1900	_sp += nargs;
duke@435	1901	do_null_check(receiver, T_OBJECT);
duke@435	1902	_sp -= nargs;
duke@435	1903	if (stopped()) {
duke@435	1904	return true;
duke@435	1905	}
duke@435	1906	// Heap pointers get a null-check from the interpreter,
duke@435	1907	// as a courtesy. However, this is not guaranteed by Unsafe,
duke@435	1908	// and it is not possible to fully distinguish unintended nulls
duke@435	1909	// from intended ones in this API.
duke@435	1910
duke@435	1911	if (is_volatile) {
duke@435	1912	// We need to emit leading and trailing CPU membars (see below) in
duke@435	1913	// addition to memory membars when is_volatile. This is a little
duke@435	1914	// too strong, but avoids the need to insert per-alias-type
duke@435	1915	// volatile membars (for stores; compare Parse::do_put_xxx), which
duke@435	1916	// we cannot do effctively here because we probably only have a
duke@435	1917	// rough approximation of type.
duke@435	1918	need_mem_bar = true;
duke@435	1919	// For Stores, place a memory ordering barrier now.
duke@435	1920	if (is_store)
duke@435	1921	insert_mem_bar(Op_MemBarRelease);
duke@435	1922	}
duke@435	1923
duke@435	1924	// Memory barrier to prevent normal and 'unsafe' accesses from
duke@435	1925	// bypassing each other. Happens after null checks, so the
duke@435	1926	// exception paths do not take memory state from the memory barrier,
duke@435	1927	// so there's no problems making a strong assert about mixing users
duke@435	1928	// of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
duke@435	1929	// around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
duke@435	1930	if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
duke@435	1931
duke@435	1932	if (!is_store) {
duke@435	1933	Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
duke@435	1934	// load value and push onto stack
duke@435	1935	switch (type) {
duke@435	1936	case T_BOOLEAN:
duke@435	1937	case T_CHAR:
duke@435	1938	case T_BYTE:
duke@435	1939	case T_SHORT:
duke@435	1940	case T_INT:
duke@435	1941	case T_FLOAT:
duke@435	1942	case T_OBJECT:
duke@435	1943	push( p );
duke@435	1944	break;
duke@435	1945	case T_ADDRESS:
duke@435	1946	// Cast to an int type.
duke@435	1947	p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
duke@435	1948	p = ConvX2L(p);
duke@435	1949	push_pair(p);
duke@435	1950	break;
duke@435	1951	case T_DOUBLE:
duke@435	1952	case T_LONG:
duke@435	1953	push_pair( p );
duke@435	1954	break;
duke@435	1955	default: ShouldNotReachHere();
duke@435	1956	}
duke@435	1957	} else {
duke@435	1958	// place effect of store into memory
duke@435	1959	switch (type) {
duke@435	1960	case T_DOUBLE:
duke@435	1961	val = dstore_rounding(val);
duke@435	1962	break;
duke@435	1963	case T_ADDRESS:
duke@435	1964	// Repackage the long as a pointer.
duke@435	1965	val = ConvL2X(val);
duke@435	1966	val = _gvn.transform( new (C, 2) CastX2PNode(val) );
duke@435	1967	break;
duke@435	1968	}
duke@435	1969
duke@435	1970	if (type != T_OBJECT ) {
duke@435	1971	(void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
duke@435	1972	} else {
duke@435	1973	// Possibly an oop being stored to Java heap or native memory
duke@435	1974	if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
duke@435	1975	// oop to Java heap.
duke@435	1976	(void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type);
duke@435	1977	} else {
duke@435	1978
duke@435	1979	// We can't tell at compile time if we are storing in the Java heap or outside
duke@435	1980	// of it. So we need to emit code to conditionally do the proper type of
duke@435	1981	// store.
duke@435	1982
duke@435	1983	IdealKit kit(gvn(), control(), merged_memory());
duke@435	1984	kit.declares_done();
duke@435	1985	// QQQ who knows what probability is here??
duke@435	1986	kit.if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
duke@435	1987	(void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type);
duke@435	1988	} kit.else_(); {
duke@435	1989	(void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
duke@435	1990	} kit.end_if();
duke@435	1991	}
duke@435	1992	}
duke@435	1993	}
duke@435	1994
duke@435	1995	if (is_volatile) {
duke@435	1996	if (!is_store)
duke@435	1997	insert_mem_bar(Op_MemBarAcquire);
duke@435	1998	else
duke@435	1999	insert_mem_bar(Op_MemBarVolatile);
duke@435	2000	}
duke@435	2001
duke@435	2002	if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2003
duke@435	2004	return true;
duke@435	2005	}
duke@435	2006
duke@435	2007	//----------------------------inline_unsafe_prefetch----------------------------
duke@435	2008
duke@435	2009	bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
duke@435	2010	#ifndef PRODUCT
duke@435	2011	{
duke@435	2012	ResourceMark rm;
duke@435	2013	// Check the signatures.
duke@435	2014	ciSignature* sig = signature();
duke@435	2015	#ifdef ASSERT
duke@435	2016	// Object getObject(Object base, int/long offset), etc.
duke@435	2017	BasicType rtype = sig->return_type()->basic_type();
duke@435	2018	if (!is_native_ptr) {
duke@435	2019	assert(sig->count() == 2, "oop prefetch has 2 arguments");
duke@435	2020	assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
duke@435	2021	assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
duke@435	2022	} else {
duke@435	2023	assert(sig->count() == 1, "native prefetch has 1 argument");
duke@435	2024	assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
duke@435	2025	}
duke@435	2026	#endif // ASSERT
duke@435	2027	}
duke@435	2028	#endif // !PRODUCT
duke@435	2029
duke@435	2030	C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@435	2031
duke@435	2032	// Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
duke@435	2033	int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
duke@435	2034
duke@435	2035	debug_only(int saved_sp = _sp);
duke@435	2036	_sp += nargs;
duke@435	2037
duke@435	2038	// Build address expression. See the code in inline_unsafe_access.
duke@435	2039	Node *adr;
duke@435	2040	if (!is_native_ptr) {
duke@435	2041	// The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
duke@435	2042	Node* offset = pop_pair();
duke@435	2043	// The base is either a Java object or a value produced by Unsafe.staticFieldBase
duke@435	2044	Node* base = pop();
duke@435	2045	// We currently rely on the cookies produced by Unsafe.xxxFieldOffset
duke@435	2046	// to be plain byte offsets, which are also the same as those accepted
duke@435	2047	// by oopDesc::field_base.
duke@435	2048	assert(Unsafe_field_offset_to_byte_offset(11) == 11,
duke@435	2049	"fieldOffset must be byte-scaled");
duke@435	2050	// 32-bit machines ignore the high half!
duke@435	2051	offset = ConvL2X(offset);
duke@435	2052	adr = make_unsafe_address(base, offset);
duke@435	2053	} else {
duke@435	2054	Node* ptr = pop_pair();
duke@435	2055	// Adjust Java long to machine word:
duke@435	2056	ptr = ConvL2X(ptr);
duke@435	2057	adr = make_unsafe_address(NULL, ptr);
duke@435	2058	}
duke@435	2059
duke@435	2060	if (is_static) {
duke@435	2061	assert(saved_sp == _sp, "must have correct argument count");
duke@435	2062	} else {
duke@435	2063	// Pop receiver last: it was pushed first.
duke@435	2064	Node *receiver = pop();
duke@435	2065	assert(saved_sp == _sp, "must have correct argument count");
duke@435	2066
duke@435	2067	// Null check on self without removing any arguments. The argument
duke@435	2068	// null check technically happens in the wrong place, which can lead to
duke@435	2069	// invalid stack traces when the primitive is inlined into a method
duke@435	2070	// which handles NullPointerExceptions.
duke@435	2071	_sp += nargs;
duke@435	2072	do_null_check(receiver, T_OBJECT);
duke@435	2073	_sp -= nargs;
duke@435	2074	if (stopped()) {
duke@435	2075	return true;
duke@435	2076	}
duke@435	2077	}
duke@435	2078
duke@435	2079	// Generate the read or write prefetch
duke@435	2080	Node *prefetch;
duke@435	2081	if (is_store) {
duke@435	2082	prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
duke@435	2083	} else {
duke@435	2084	prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
duke@435	2085	}
duke@435	2086	prefetch->init_req(0, control());
duke@435	2087	set_i_o(_gvn.transform(prefetch));
duke@435	2088
duke@435	2089	return true;
duke@435	2090	}
duke@435	2091
duke@435	2092	//----------------------------inline_unsafe_CAS----------------------------
duke@435	2093
duke@435	2094	bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
duke@435	2095	// This basic scheme here is the same as inline_unsafe_access, but
duke@435	2096	// differs in enough details that combining them would make the code
duke@435	2097	// overly confusing. (This is a true fact! I originally combined
duke@435	2098	// them, but even I was confused by it!) As much code/comments as
duke@435	2099	// possible are retained from inline_unsafe_access though to make
duke@435	2100	// the correspondances clearer. - dl
duke@435	2101
duke@435	2102	if (callee()->is_static()) return false; // caller must have the capability!
duke@435	2103
duke@435	2104	#ifndef PRODUCT
duke@435	2105	{
duke@435	2106	ResourceMark rm;
duke@435	2107	// Check the signatures.
duke@435	2108	ciSignature* sig = signature();
duke@435	2109	#ifdef ASSERT
duke@435	2110	BasicType rtype = sig->return_type()->basic_type();
duke@435	2111	assert(rtype == T_BOOLEAN, "CAS must return boolean");
duke@435	2112	assert(sig->count() == 4, "CAS has 4 arguments");
duke@435	2113	assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
duke@435	2114	assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
duke@435	2115	#endif // ASSERT
duke@435	2116	}
duke@435	2117	#endif //PRODUCT
duke@435	2118
duke@435	2119	// number of stack slots per value argument (1 or 2)
duke@435	2120	int type_words = type2size[type];
duke@435	2121
duke@435	2122	// Cannot inline wide CAS on machines that don't support it natively
kvn@464	2123	if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
duke@435	2124	return false;
duke@435	2125
duke@435	2126	C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@435	2127
duke@435	2128	// Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
duke@435	2129	int nargs = 1 + 1 + 2 + type_words + type_words;
duke@435	2130
duke@435	2131	// pop arguments: newval, oldval, offset, base, and receiver
duke@435	2132	debug_only(int saved_sp = _sp);
duke@435	2133	_sp += nargs;
duke@435	2134	Node* newval = (type_words == 1) ? pop() : pop_pair();
duke@435	2135	Node* oldval = (type_words == 1) ? pop() : pop_pair();
duke@435	2136	Node *offset = pop_pair();
duke@435	2137	Node *base = pop();
duke@435	2138	Node *receiver = pop();
duke@435	2139	assert(saved_sp == _sp, "must have correct argument count");
duke@435	2140
duke@435	2141	// Null check receiver.
duke@435	2142	_sp += nargs;
duke@435	2143	do_null_check(receiver, T_OBJECT);
duke@435	2144	_sp -= nargs;
duke@435	2145	if (stopped()) {
duke@435	2146	return true;
duke@435	2147	}
duke@435	2148
duke@435	2149	// Build field offset expression.
duke@435	2150	// We currently rely on the cookies produced by Unsafe.xxxFieldOffset
duke@435	2151	// to be plain byte offsets, which are also the same as those accepted
duke@435	2152	// by oopDesc::field_base.
duke@435	2153	assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
duke@435	2154	// 32-bit machines ignore the high half of long offsets
duke@435	2155	offset = ConvL2X(offset);
duke@435	2156	Node* adr = make_unsafe_address(base, offset);
duke@435	2157	const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
duke@435	2158
duke@435	2159	// (Unlike inline_unsafe_access, there seems no point in trying
duke@435	2160	// to refine types. Just use the coarse types here.
duke@435	2161	const Type *value_type = Type::get_const_basic_type(type);
duke@435	2162	Compile::AliasType* alias_type = C->alias_type(adr_type);
duke@435	2163	assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
duke@435	2164	int alias_idx = C->get_alias_index(adr_type);
duke@435	2165
duke@435	2166	// Memory-model-wise, a CAS acts like a little synchronized block,
duke@435	2167	// so needs barriers on each side. These don't't translate into
duke@435	2168	// actual barriers on most machines, but we still need rest of
duke@435	2169	// compiler to respect ordering.
duke@435	2170
duke@435	2171	insert_mem_bar(Op_MemBarRelease);
duke@435	2172	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2173
duke@435	2174	// 4984716: MemBars must be inserted before this
duke@435	2175	// memory node in order to avoid a false
duke@435	2176	// dependency which will confuse the scheduler.
duke@435	2177	Node *mem = memory(alias_idx);
duke@435	2178
duke@435	2179	// For now, we handle only those cases that actually exist: ints,
duke@435	2180	// longs, and Object. Adding others should be straightforward.
duke@435	2181	Node* cas;
duke@435	2182	switch(type) {
duke@435	2183	case T_INT:
duke@435	2184	cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
duke@435	2185	break;
duke@435	2186	case T_LONG:
duke@435	2187	cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
duke@435	2188	break;
duke@435	2189	case T_OBJECT:
coleenp@548	2190	// reference stores need a store barrier.
duke@435	2191	// (They don't if CAS fails, but it isn't worth checking.)
duke@435	2192	pre_barrier(control(), base, adr, alias_idx, newval, value_type, T_OBJECT);
coleenp@548	2193	#ifdef _LP64
kvn@598	2194	if (adr->bottom_type()->is_ptr_to_narrowoop()) {
coleenp@548	2195	cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
coleenp@548	2196	EncodePNode::encode(&_gvn, newval),
coleenp@548	2197	EncodePNode::encode(&_gvn, oldval)));
coleenp@548	2198	} else
coleenp@548	2199	#endif
coleenp@548	2200	{
coleenp@548	2201	cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
coleenp@548	2202	}
duke@435	2203	post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
duke@435	2204	break;
duke@435	2205	default:
duke@435	2206	ShouldNotReachHere();
duke@435	2207	break;
duke@435	2208	}
duke@435	2209
duke@435	2210	// SCMemProjNodes represent the memory state of CAS. Their main
duke@435	2211	// role is to prevent CAS nodes from being optimized away when their
duke@435	2212	// results aren't used.
duke@435	2213	Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
duke@435	2214	set_memory(proj, alias_idx);
duke@435	2215
duke@435	2216	// Add the trailing membar surrounding the access
duke@435	2217	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2218	insert_mem_bar(Op_MemBarAcquire);
duke@435	2219
duke@435	2220	push(cas);
duke@435	2221	return true;
duke@435	2222	}
duke@435	2223
duke@435	2224	bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
duke@435	2225	// This is another variant of inline_unsafe_access, differing in
duke@435	2226	// that it always issues store-store ("release") barrier and ensures
duke@435	2227	// store-atomicity (which only matters for "long").
duke@435	2228
duke@435	2229	if (callee()->is_static()) return false; // caller must have the capability!
duke@435	2230
duke@435	2231	#ifndef PRODUCT
duke@435	2232	{
duke@435	2233	ResourceMark rm;
duke@435	2234	// Check the signatures.
duke@435	2235	ciSignature* sig = signature();
duke@435	2236	#ifdef ASSERT
duke@435	2237	BasicType rtype = sig->return_type()->basic_type();
duke@435	2238	assert(rtype == T_VOID, "must return void");
duke@435	2239	assert(sig->count() == 3, "has 3 arguments");
duke@435	2240	assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
duke@435	2241	assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
duke@435	2242	#endif // ASSERT
duke@435	2243	}
duke@435	2244	#endif //PRODUCT
duke@435	2245
duke@435	2246	// number of stack slots per value argument (1 or 2)
duke@435	2247	int type_words = type2size[type];
duke@435	2248
duke@435	2249	C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@435	2250
duke@435	2251	// Argument words: "this" plus oop plus offset plus value;
duke@435	2252	int nargs = 1 + 1 + 2 + type_words;
duke@435	2253
duke@435	2254	// pop arguments: val, offset, base, and receiver
duke@435	2255	debug_only(int saved_sp = _sp);
duke@435	2256	_sp += nargs;
duke@435	2257	Node* val = (type_words == 1) ? pop() : pop_pair();
duke@435	2258	Node *offset = pop_pair();
duke@435	2259	Node *base = pop();
duke@435	2260	Node *receiver = pop();
duke@435	2261	assert(saved_sp == _sp, "must have correct argument count");
duke@435	2262
duke@435	2263	// Null check receiver.
duke@435	2264	_sp += nargs;
duke@435	2265	do_null_check(receiver, T_OBJECT);
duke@435	2266	_sp -= nargs;
duke@435	2267	if (stopped()) {
duke@435	2268	return true;
duke@435	2269	}
duke@435	2270
duke@435	2271	// Build field offset expression.
duke@435	2272	assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
duke@435	2273	// 32-bit machines ignore the high half of long offsets
duke@435	2274	offset = ConvL2X(offset);
duke@435	2275	Node* adr = make_unsafe_address(base, offset);
duke@435	2276	const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
duke@435	2277	const Type *value_type = Type::get_const_basic_type(type);
duke@435	2278	Compile::AliasType* alias_type = C->alias_type(adr_type);
duke@435	2279
duke@435	2280	insert_mem_bar(Op_MemBarRelease);
duke@435	2281	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2282	// Ensure that the store is atomic for longs:
duke@435	2283	bool require_atomic_access = true;
duke@435	2284	Node* store;
duke@435	2285	if (type == T_OBJECT) // reference stores need a store barrier.
duke@435	2286	store = store_oop_to_unknown(control(), base, adr, adr_type, val, value_type, type);
duke@435	2287	else {
duke@435	2288	store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
duke@435	2289	}
duke@435	2290	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2291	return true;
duke@435	2292	}
duke@435	2293
duke@435	2294	bool LibraryCallKit::inline_unsafe_allocate() {
duke@435	2295	if (callee()->is_static()) return false; // caller must have the capability!
duke@435	2296	int nargs = 1 + 1;
duke@435	2297	assert(signature()->size() == nargs-1, "alloc has 1 argument");
duke@435	2298	null_check_receiver(callee()); // check then ignore argument(0)
duke@435	2299	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2300	Node* cls = do_null_check(argument(1), T_OBJECT);
duke@435	2301	_sp -= nargs;
duke@435	2302	if (stopped()) return true;
duke@435	2303
duke@435	2304	Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
duke@435	2305	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2306	kls = do_null_check(kls, T_OBJECT);
duke@435	2307	_sp -= nargs;
duke@435	2308	if (stopped()) return true; // argument was like int.class
duke@435	2309
duke@435	2310	// Note: The argument might still be an illegal value like
duke@435	2311	// Serializable.class or Object[].class. The runtime will handle it.
duke@435	2312	// But we must make an explicit check for initialization.
duke@435	2313	Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
duke@435	2314	Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
duke@435	2315	Node* bits = intcon(instanceKlass::fully_initialized);
duke@435	2316	Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
duke@435	2317	// The 'test' is non-zero if we need to take a slow path.
duke@435	2318
duke@435	2319	Node* obj = new_instance(kls, test);
duke@435	2320	push(obj);
duke@435	2321
duke@435	2322	return true;
duke@435	2323	}
duke@435	2324
duke@435	2325	//------------------------inline_native_time_funcs--------------
duke@435	2326	// inline code for System.currentTimeMillis() and System.nanoTime()
duke@435	2327	// these have the same type and signature
duke@435	2328	bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
duke@435	2329	address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
duke@435	2330	CAST_FROM_FN_PTR(address, os::javaTimeMillis);
duke@435	2331	const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
duke@435	2332	const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
duke@435	2333	const TypePtr* no_memory_effects = NULL;
duke@435	2334	Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
duke@435	2335	Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
duke@435	2336	#ifdef ASSERT
duke@435	2337	Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
duke@435	2338	assert(value_top == top(), "second value must be top");
duke@435	2339	#endif
duke@435	2340	push_pair(value);
duke@435	2341	return true;
duke@435	2342	}
duke@435	2343
duke@435	2344	//------------------------inline_native_currentThread------------------
duke@435	2345	bool LibraryCallKit::inline_native_currentThread() {
duke@435	2346	Node* junk = NULL;
duke@435	2347	push(generate_current_thread(junk));
duke@435	2348	return true;
duke@435	2349	}
duke@435	2350
duke@435	2351	//------------------------inline_native_isInterrupted------------------
duke@435	2352	bool LibraryCallKit::inline_native_isInterrupted() {
duke@435	2353	const int nargs = 1+1; // receiver + boolean
duke@435	2354	assert(nargs == arg_size(), "sanity");
duke@435	2355	// Add a fast path to t.isInterrupted(clear_int):
duke@435	2356	// (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
duke@435	2357	// ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
duke@435	2358	// So, in the common case that the interrupt bit is false,
duke@435	2359	// we avoid making a call into the VM. Even if the interrupt bit
duke@435	2360	// is true, if the clear_int argument is false, we avoid the VM call.
duke@435	2361	// However, if the receiver is not currentThread, we must call the VM,
duke@435	2362	// because there must be some locking done around the operation.
duke@435	2363
duke@435	2364	// We only go to the fast case code if we pass two guards.
duke@435	2365	// Paths which do not pass are accumulated in the slow_region.
duke@435	2366	RegionNode* slow_region = new (C, 1) RegionNode(1);
duke@435	2367	record_for_igvn(slow_region);
duke@435	2368	RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
duke@435	2369	PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
duke@435	2370	enum { no_int_result_path = 1,
duke@435	2371	no_clear_result_path = 2,
duke@435	2372	slow_result_path = 3
duke@435	2373	};
duke@435	2374
duke@435	2375	// (a) Receiving thread must be the current thread.
duke@435	2376	Node* rec_thr = argument(0);
duke@435	2377	Node* tls_ptr = NULL;
duke@435	2378	Node* cur_thr = generate_current_thread(tls_ptr);
duke@435	2379	Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
duke@435	2380	Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
duke@435	2381
duke@435	2382	bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
duke@435	2383	if (!known_current_thread)
duke@435	2384	generate_slow_guard(bol_thr, slow_region);
duke@435	2385
duke@435	2386	// (b) Interrupt bit on TLS must be false.
duke@435	2387	Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
duke@435	2388	Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
duke@435	2389	p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
duke@435	2390	Node* int_bit = make_load(NULL, p, TypeInt::BOOL, T_INT);
duke@435	2391	Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
duke@435	2392	Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
duke@435	2393
duke@435	2394	IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
duke@435	2395
duke@435	2396	// First fast path: if (!TLS._interrupted) return false;
duke@435	2397	Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
duke@435	2398	result_rgn->init_req(no_int_result_path, false_bit);
duke@435	2399	result_val->init_req(no_int_result_path, intcon(0));
duke@435	2400
duke@435	2401	// drop through to next case
duke@435	2402	set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
duke@435	2403
duke@435	2404	// (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
duke@435	2405	Node* clr_arg = argument(1);
duke@435	2406	Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
duke@435	2407	Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
duke@435	2408	IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
duke@435	2409
duke@435	2410	// Second fast path: ... else if (!clear_int) return true;
duke@435	2411	Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
duke@435	2412	result_rgn->init_req(no_clear_result_path, false_arg);
duke@435	2413	result_val->init_req(no_clear_result_path, intcon(1));
duke@435	2414
duke@435	2415	// drop through to next case
duke@435	2416	set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
duke@435	2417
duke@435	2418	// (d) Otherwise, go to the slow path.
duke@435	2419	slow_region->add_req(control());
duke@435	2420	set_control( _gvn.transform(slow_region) );
duke@435	2421
duke@435	2422	if (stopped()) {
duke@435	2423	// There is no slow path.
duke@435	2424	result_rgn->init_req(slow_result_path, top());
duke@435	2425	result_val->init_req(slow_result_path, top());
duke@435	2426	} else {
duke@435	2427	// non-virtual because it is a private non-static
duke@435	2428	CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
duke@435	2429
duke@435	2430	Node* slow_val = set_results_for_java_call(slow_call);
duke@435	2431	// this->control() comes from set_results_for_java_call
duke@435	2432
duke@435	2433	// If we know that the result of the slow call will be true, tell the optimizer!
duke@435	2434	if (known_current_thread) slow_val = intcon(1);
duke@435	2435
duke@435	2436	Node* fast_io = slow_call->in(TypeFunc::I_O);
duke@435	2437	Node* fast_mem = slow_call->in(TypeFunc::Memory);
duke@435	2438	// These two phis are pre-filled with copies of of the fast IO and Memory
duke@435	2439	Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
duke@435	2440	Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
duke@435	2441
duke@435	2442	result_rgn->init_req(slow_result_path, control());
duke@435	2443	io_phi ->init_req(slow_result_path, i_o());
duke@435	2444	mem_phi ->init_req(slow_result_path, reset_memory());
duke@435	2445	result_val->init_req(slow_result_path, slow_val);
duke@435	2446
duke@435	2447	set_all_memory( _gvn.transform(mem_phi) );
duke@435	2448	set_i_o( _gvn.transform(io_phi) );
duke@435	2449	}
duke@435	2450
duke@435	2451	push_result(result_rgn, result_val);
duke@435	2452	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	2453
duke@435	2454	return true;
duke@435	2455	}
duke@435	2456
duke@435	2457	//---------------------------load_mirror_from_klass----------------------------
duke@435	2458	// Given a klass oop, load its java mirror (a java.lang.Class oop).
duke@435	2459	Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
duke@435	2460	Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
duke@435	2461	return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
duke@435	2462	}
duke@435	2463
duke@435	2464	//-----------------------load_klass_from_mirror_common-------------------------
duke@435	2465	// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
duke@435	2466	// Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
duke@435	2467	// and branch to the given path on the region.
duke@435	2468	// If never_see_null, take an uncommon trap on null, so we can optimistically
duke@435	2469	// compile for the non-null case.
duke@435	2470	// If the region is NULL, force never_see_null = true.
duke@435	2471	Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
duke@435	2472	bool never_see_null,
duke@435	2473	int nargs,
duke@435	2474	RegionNode* region,
duke@435	2475	int null_path,
duke@435	2476	int offset) {
duke@435	2477	if (region == NULL) never_see_null = true;
duke@435	2478	Node* p = basic_plus_adr(mirror, offset);
duke@435	2479	const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
kvn@599	2480	Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
duke@435	2481	_sp += nargs; // any deopt will start just before call to enclosing method
duke@435	2482	Node* null_ctl = top();
duke@435	2483	kls = null_check_oop(kls, &null_ctl, never_see_null);
duke@435	2484	if (region != NULL) {
duke@435	2485	// Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
duke@435	2486	region->init_req(null_path, null_ctl);
duke@435	2487	} else {
duke@435	2488	assert(null_ctl == top(), "no loose ends");
duke@435	2489	}
duke@435	2490	_sp -= nargs;
duke@435	2491	return kls;
duke@435	2492	}
duke@435	2493
duke@435	2494	//--------------------(inline_native_Class_query helpers)---------------------
duke@435	2495	// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
duke@435	2496	// Fall through if (mods & mask) == bits, take the guard otherwise.
duke@435	2497	Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
duke@435	2498	// Branch around if the given klass has the given modifier bit set.
duke@435	2499	// Like generate_guard, adds a new path onto the region.
duke@435	2500	Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
duke@435	2501	Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
duke@435	2502	Node* mask = intcon(modifier_mask);
duke@435	2503	Node* bits = intcon(modifier_bits);
duke@435	2504	Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
duke@435	2505	Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
duke@435	2506	Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
duke@435	2507	return generate_fair_guard(bol, region);
duke@435	2508	}
duke@435	2509	Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
duke@435	2510	return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
duke@435	2511	}
duke@435	2512
duke@435	2513	//-------------------------inline_native_Class_query-------------------
duke@435	2514	bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
duke@435	2515	int nargs = 1+0; // just the Class mirror, in most cases
duke@435	2516	const Type* return_type = TypeInt::BOOL;
duke@435	2517	Node* prim_return_value = top(); // what happens if it's a primitive class?
duke@435	2518	bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
duke@435	2519	bool expect_prim = false; // most of these guys expect to work on refs
duke@435	2520
duke@435	2521	enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
duke@435	2522
duke@435	2523	switch (id) {
duke@435	2524	case vmIntrinsics::_isInstance:
duke@435	2525	nargs = 1+1; // the Class mirror, plus the object getting queried about
duke@435	2526	// nothing is an instance of a primitive type
duke@435	2527	prim_return_value = intcon(0);
duke@435	2528	break;
duke@435	2529	case vmIntrinsics::_getModifiers:
duke@435	2530	prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
duke@435	2531	assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
duke@435	2532	return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
duke@435	2533	break;
duke@435	2534	case vmIntrinsics::_isInterface:
duke@435	2535	prim_return_value = intcon(0);
duke@435	2536	break;
duke@435	2537	case vmIntrinsics::_isArray:
duke@435	2538	prim_return_value = intcon(0);
duke@435	2539	expect_prim = true; // cf. ObjectStreamClass.getClassSignature
duke@435	2540	break;
duke@435	2541	case vmIntrinsics::_isPrimitive:
duke@435	2542	prim_return_value = intcon(1);
duke@435	2543	expect_prim = true; // obviously
duke@435	2544	break;
duke@435	2545	case vmIntrinsics::_getSuperclass:
duke@435	2546	prim_return_value = null();
duke@435	2547	return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
duke@435	2548	break;
duke@435	2549	case vmIntrinsics::_getComponentType:
duke@435	2550	prim_return_value = null();
duke@435	2551	return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
duke@435	2552	break;
duke@435	2553	case vmIntrinsics::_getClassAccessFlags:
duke@435	2554	prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
duke@435	2555	return_type = TypeInt::INT; // not bool! 6297094
duke@435	2556	break;
duke@435	2557	default:
duke@435	2558	ShouldNotReachHere();
duke@435	2559	}
duke@435	2560
duke@435	2561	Node* mirror = argument(0);
duke@435	2562	Node* obj = (nargs <= 1)? top(): argument(1);
duke@435	2563
duke@435	2564	const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
duke@435	2565	if (mirror_con == NULL) return false; // cannot happen?
duke@435	2566
duke@435	2567	#ifndef PRODUCT
duke@435	2568	if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
duke@435	2569	ciType* k = mirror_con->java_mirror_type();
duke@435	2570	if (k) {
duke@435	2571	tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
duke@435	2572	k->print_name();
duke@435	2573	tty->cr();
duke@435	2574	}
duke@435	2575	}
duke@435	2576	#endif
duke@435	2577
duke@435	2578	// Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
duke@435	2579	RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	2580	record_for_igvn(region);
duke@435	2581	PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
duke@435	2582
duke@435	2583	// The mirror will never be null of Reflection.getClassAccessFlags, however
duke@435	2584	// it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
duke@435	2585	// if it is. See bug 4774291.
duke@435	2586
duke@435	2587	// For Reflection.getClassAccessFlags(), the null check occurs in
duke@435	2588	// the wrong place; see inline_unsafe_access(), above, for a similar
duke@435	2589	// situation.
duke@435	2590	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2591	mirror = do_null_check(mirror, T_OBJECT);
duke@435	2592	_sp -= nargs;
duke@435	2593	// If mirror or obj is dead, only null-path is taken.
duke@435	2594	if (stopped()) return true;
duke@435	2595
duke@435	2596	if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
duke@435	2597
duke@435	2598	// Now load the mirror's klass metaobject, and null-check it.
duke@435	2599	// Side-effects region with the control path if the klass is null.
duke@435	2600	Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
duke@435	2601	region, _prim_path);
duke@435	2602	// If kls is null, we have a primitive mirror.
duke@435	2603	phi->init_req(_prim_path, prim_return_value);
duke@435	2604	if (stopped()) { push_result(region, phi); return true; }
duke@435	2605
duke@435	2606	Node* p; // handy temp
duke@435	2607	Node* null_ctl;
duke@435	2608
duke@435	2609	// Now that we have the non-null klass, we can perform the real query.
duke@435	2610	// For constant classes, the query will constant-fold in LoadNode::Value.
duke@435	2611	Node* query_value = top();
duke@435	2612	switch (id) {
duke@435	2613	case vmIntrinsics::_isInstance:
duke@435	2614	// nothing is an instance of a primitive type
duke@435	2615	query_value = gen_instanceof(obj, kls);
duke@435	2616	break;
duke@435	2617
duke@435	2618	case vmIntrinsics::_getModifiers:
duke@435	2619	p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
duke@435	2620	query_value = make_load(NULL, p, TypeInt::INT, T_INT);
duke@435	2621	break;
duke@435	2622
duke@435	2623	case vmIntrinsics::_isInterface:
duke@435	2624	// (To verify this code sequence, check the asserts in JVM_IsInterface.)
duke@435	2625	if (generate_interface_guard(kls, region) != NULL)
duke@435	2626	// A guard was added. If the guard is taken, it was an interface.
duke@435	2627	phi->add_req(intcon(1));
duke@435	2628	// If we fall through, it's a plain class.
duke@435	2629	query_value = intcon(0);
duke@435	2630	break;
duke@435	2631
duke@435	2632	case vmIntrinsics::_isArray:
duke@435	2633	// (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
duke@435	2634	if (generate_array_guard(kls, region) != NULL)
duke@435	2635	// A guard was added. If the guard is taken, it was an array.
duke@435	2636	phi->add_req(intcon(1));
duke@435	2637	// If we fall through, it's a plain class.
duke@435	2638	query_value = intcon(0);
duke@435	2639	break;
duke@435	2640
duke@435	2641	case vmIntrinsics::_isPrimitive:
duke@435	2642	query_value = intcon(0); // "normal" path produces false
duke@435	2643	break;
duke@435	2644
duke@435	2645	case vmIntrinsics::_getSuperclass:
duke@435	2646	// The rules here are somewhat unfortunate, but we can still do better
duke@435	2647	// with random logic than with a JNI call.
duke@435	2648	// Interfaces store null or Object as _super, but must report null.
duke@435	2649	// Arrays store an intermediate super as _super, but must report Object.
duke@435	2650	// Other types can report the actual _super.
duke@435	2651	// (To verify this code sequence, check the asserts in JVM_IsInterface.)
duke@435	2652	if (generate_interface_guard(kls, region) != NULL)
duke@435	2653	// A guard was added. If the guard is taken, it was an interface.
duke@435	2654	phi->add_req(null());
duke@435	2655	if (generate_array_guard(kls, region) != NULL)
duke@435	2656	// A guard was added. If the guard is taken, it was an array.
duke@435	2657	phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
duke@435	2658	// If we fall through, it's a plain class. Get its _super.
duke@435	2659	p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
kvn@599	2660	kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
duke@435	2661	null_ctl = top();
duke@435	2662	kls = null_check_oop(kls, &null_ctl);
duke@435	2663	if (null_ctl != top()) {
duke@435	2664	// If the guard is taken, Object.superClass is null (both klass and mirror).
duke@435	2665	region->add_req(null_ctl);
duke@435	2666	phi ->add_req(null());
duke@435	2667	}
duke@435	2668	if (!stopped()) {
duke@435	2669	query_value = load_mirror_from_klass(kls);
duke@435	2670	}
duke@435	2671	break;
duke@435	2672
duke@435	2673	case vmIntrinsics::_getComponentType:
duke@435	2674	if (generate_array_guard(kls, region) != NULL) {
duke@435	2675	// Be sure to pin the oop load to the guard edge just created:
duke@435	2676	Node* is_array_ctrl = region->in(region->req()-1);
duke@435	2677	Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
duke@435	2678	Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
duke@435	2679	phi->add_req(cmo);
duke@435	2680	}
duke@435	2681	query_value = null(); // non-array case is null
duke@435	2682	break;
duke@435	2683
duke@435	2684	case vmIntrinsics::_getClassAccessFlags:
duke@435	2685	p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
duke@435	2686	query_value = make_load(NULL, p, TypeInt::INT, T_INT);
duke@435	2687	break;
duke@435	2688
duke@435	2689	default:
duke@435	2690	ShouldNotReachHere();
duke@435	2691	}
duke@435	2692
duke@435	2693	// Fall-through is the normal case of a query to a real class.
duke@435	2694	phi->init_req(1, query_value);
duke@435	2695	region->init_req(1, control());
duke@435	2696
duke@435	2697	push_result(region, phi);
duke@435	2698	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	2699
duke@435	2700	return true;
duke@435	2701	}
duke@435	2702
duke@435	2703	//--------------------------inline_native_subtype_check------------------------
duke@435	2704	// This intrinsic takes the JNI calls out of the heart of
duke@435	2705	// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
duke@435	2706	bool LibraryCallKit::inline_native_subtype_check() {
duke@435	2707	int nargs = 1+1; // the Class mirror, plus the other class getting examined
duke@435	2708
duke@435	2709	// Pull both arguments off the stack.
duke@435	2710	Node* args[2]; // two java.lang.Class mirrors: superc, subc
duke@435	2711	args[0] = argument(0);
duke@435	2712	args[1] = argument(1);
duke@435	2713	Node* klasses[2]; // corresponding Klasses: superk, subk
duke@435	2714	klasses[0] = klasses[1] = top();
duke@435	2715
duke@435	2716	enum {
duke@435	2717	// A full decision tree on {superc is prim, subc is prim}:
duke@435	2718	_prim_0_path = 1, // {P,N} => false
duke@435	2719	// {P,P} & superc!=subc => false
duke@435	2720	_prim_same_path, // {P,P} & superc==subc => true
duke@435	2721	_prim_1_path, // {N,P} => false
duke@435	2722	_ref_subtype_path, // {N,N} & subtype check wins => true
duke@435	2723	_both_ref_path, // {N,N} & subtype check loses => false
duke@435	2724	PATH_LIMIT
duke@435	2725	};
duke@435	2726
duke@435	2727	RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	2728	Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
duke@435	2729	record_for_igvn(region);
duke@435	2730
duke@435	2731	const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
duke@435	2732	const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
duke@435	2733	int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
duke@435	2734
duke@435	2735	// First null-check both mirrors and load each mirror's klass metaobject.
duke@435	2736	int which_arg;
duke@435	2737	for (which_arg = 0; which_arg <= 1; which_arg++) {
duke@435	2738	Node* arg = args[which_arg];
duke@435	2739	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2740	arg = do_null_check(arg, T_OBJECT);
duke@435	2741	_sp -= nargs;
duke@435	2742	if (stopped()) break;
duke@435	2743	args[which_arg] = _gvn.transform(arg);
duke@435	2744
duke@435	2745	Node* p = basic_plus_adr(arg, class_klass_offset);
kvn@599	2746	Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
duke@435	2747	klasses[which_arg] = _gvn.transform(kls);
duke@435	2748	}
duke@435	2749
duke@435	2750	// Having loaded both klasses, test each for null.
duke@435	2751	bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
duke@435	2752	for (which_arg = 0; which_arg <= 1; which_arg++) {
duke@435	2753	Node* kls = klasses[which_arg];
duke@435	2754	Node* null_ctl = top();
duke@435	2755	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2756	kls = null_check_oop(kls, &null_ctl, never_see_null);
duke@435	2757	_sp -= nargs;
duke@435	2758	int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
duke@435	2759	region->init_req(prim_path, null_ctl);
duke@435	2760	if (stopped()) break;
duke@435	2761	klasses[which_arg] = kls;
duke@435	2762	}
duke@435	2763
duke@435	2764	if (!stopped()) {
duke@435	2765	// now we have two reference types, in klasses[0..1]
duke@435	2766	Node* subk = klasses[1]; // the argument to isAssignableFrom
duke@435	2767	Node* superk = klasses[0]; // the receiver
duke@435	2768	region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
duke@435	2769	// now we have a successful reference subtype check
duke@435	2770	region->set_req(_ref_subtype_path, control());
duke@435	2771	}
duke@435	2772
duke@435	2773	// If both operands are primitive (both klasses null), then
duke@435	2774	// we must return true when they are identical primitives.
duke@435	2775	// It is convenient to test this after the first null klass check.
duke@435	2776	set_control(region->in(_prim_0_path)); // go back to first null check
duke@435	2777	if (!stopped()) {
duke@435	2778	// Since superc is primitive, make a guard for the superc==subc case.
duke@435	2779	Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
duke@435	2780	Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
duke@435	2781	generate_guard(bol_eq, region, PROB_FAIR);
duke@435	2782	if (region->req() == PATH_LIMIT+1) {
duke@435	2783	// A guard was added. If the added guard is taken, superc==subc.
duke@435	2784	region->swap_edges(PATH_LIMIT, _prim_same_path);
duke@435	2785	region->del_req(PATH_LIMIT);
duke@435	2786	}
duke@435	2787	region->set_req(_prim_0_path, control()); // Not equal after all.
duke@435	2788	}
duke@435	2789
duke@435	2790	// these are the only paths that produce 'true':
duke@435	2791	phi->set_req(_prim_same_path, intcon(1));
duke@435	2792	phi->set_req(_ref_subtype_path, intcon(1));
duke@435	2793
duke@435	2794	// pull together the cases:
duke@435	2795	assert(region->req() == PATH_LIMIT, "sane region");
duke@435	2796	for (uint i = 1; i < region->req(); i++) {
duke@435	2797	Node* ctl = region->in(i);
duke@435	2798	if (ctl == NULL || ctl == top()) {
duke@435	2799	region->set_req(i, top());
duke@435	2800	phi ->set_req(i, top());
duke@435	2801	} else if (phi->in(i) == NULL) {
duke@435	2802	phi->set_req(i, intcon(0)); // all other paths produce 'false'
duke@435	2803	}
duke@435	2804	}
duke@435	2805
duke@435	2806	set_control(_gvn.transform(region));
duke@435	2807	push(_gvn.transform(phi));
duke@435	2808
duke@435	2809	return true;
duke@435	2810	}
duke@435	2811
duke@435	2812	//---------------------generate_array_guard_common------------------------
duke@435	2813	Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
duke@435	2814	bool obj_array, bool not_array) {
duke@435	2815	// If obj_array/non_array==false/false:
duke@435	2816	// Branch around if the given klass is in fact an array (either obj or prim).
duke@435	2817	// If obj_array/non_array==false/true:
duke@435	2818	// Branch around if the given klass is not an array klass of any kind.
duke@435	2819	// If obj_array/non_array==true/true:
duke@435	2820	// Branch around if the kls is not an oop array (kls is int[], String, etc.)
duke@435	2821	// If obj_array/non_array==true/false:
duke@435	2822	// Branch around if the kls is an oop array (Object[] or subtype)
duke@435	2823	//
duke@435	2824	// Like generate_guard, adds a new path onto the region.
duke@435	2825	jint layout_con = 0;
duke@435	2826	Node* layout_val = get_layout_helper(kls, layout_con);
duke@435	2827	if (layout_val == NULL) {
duke@435	2828	bool query = (obj_array
duke@435	2829	? Klass::layout_helper_is_objArray(layout_con)
duke@435	2830	: Klass::layout_helper_is_javaArray(layout_con));
duke@435	2831	if (query == not_array) {
duke@435	2832	return NULL; // never a branch
duke@435	2833	} else { // always a branch
duke@435	2834	Node* always_branch = control();
duke@435	2835	if (region != NULL)
duke@435	2836	region->add_req(always_branch);
duke@435	2837	set_control(top());
duke@435	2838	return always_branch;
duke@435	2839	}
duke@435	2840	}
duke@435	2841	// Now test the correct condition.
duke@435	2842	jint nval = (obj_array
duke@435	2843	? ((jint)Klass::_lh_array_tag_type_value
duke@435	2844	<< Klass::_lh_array_tag_shift)
duke@435	2845	: Klass::_lh_neutral_value);
duke@435	2846	Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
duke@435	2847	BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
duke@435	2848	// invert the test if we are looking for a non-array
duke@435	2849	if (not_array) btest = BoolTest(btest).negate();
duke@435	2850	Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
duke@435	2851	return generate_fair_guard(bol, region);
duke@435	2852	}
duke@435	2853
duke@435	2854
duke@435	2855	//-----------------------inline_native_newArray--------------------------
duke@435	2856	bool LibraryCallKit::inline_native_newArray() {
duke@435	2857	int nargs = 2;
duke@435	2858	Node* mirror = argument(0);
duke@435	2859	Node* count_val = argument(1);
duke@435	2860
duke@435	2861	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2862	mirror = do_null_check(mirror, T_OBJECT);
duke@435	2863	_sp -= nargs;
kvn@598	2864	// If mirror or obj is dead, only null-path is taken.
kvn@598	2865	if (stopped()) return true;
duke@435	2866
duke@435	2867	enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
duke@435	2868	RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	2869	PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
duke@435	2870	TypeInstPtr::NOTNULL);
duke@435	2871	PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
duke@435	2872	PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
duke@435	2873	TypePtr::BOTTOM);
duke@435	2874
duke@435	2875	bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
duke@435	2876	Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
duke@435	2877	nargs,
duke@435	2878	result_reg, _slow_path);
duke@435	2879	Node* normal_ctl = control();
duke@435	2880	Node* no_array_ctl = result_reg->in(_slow_path);
duke@435	2881
duke@435	2882	// Generate code for the slow case. We make a call to newArray().
duke@435	2883	set_control(no_array_ctl);
duke@435	2884	if (!stopped()) {
duke@435	2885	// Either the input type is void.class, or else the
duke@435	2886	// array klass has not yet been cached. Either the
duke@435	2887	// ensuing call will throw an exception, or else it
duke@435	2888	// will cache the array klass for next time.
duke@435	2889	PreserveJVMState pjvms(this);
duke@435	2890	CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
duke@435	2891	Node* slow_result = set_results_for_java_call(slow_call);
duke@435	2892	// this->control() comes from set_results_for_java_call
duke@435	2893	result_reg->set_req(_slow_path, control());
duke@435	2894	result_val->set_req(_slow_path, slow_result);
duke@435	2895	result_io ->set_req(_slow_path, i_o());
duke@435	2896	result_mem->set_req(_slow_path, reset_memory());
duke@435	2897	}
duke@435	2898
duke@435	2899	set_control(normal_ctl);
duke@435	2900	if (!stopped()) {
duke@435	2901	// Normal case: The array type has been cached in the java.lang.Class.
duke@435	2902	// The following call works fine even if the array type is polymorphic.
duke@435	2903	// It could be a dynamic mix of int[], boolean[], Object[], etc.
duke@435	2904	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2905	Node* obj = new_array(klass_node, count_val);
duke@435	2906	_sp -= nargs;
duke@435	2907	result_reg->init_req(_normal_path, control());
duke@435	2908	result_val->init_req(_normal_path, obj);
duke@435	2909	result_io ->init_req(_normal_path, i_o());
duke@435	2910	result_mem->init_req(_normal_path, reset_memory());
duke@435	2911	}
duke@435	2912
duke@435	2913	// Return the combined state.
duke@435	2914	set_i_o( _gvn.transform(result_io) );
duke@435	2915	set_all_memory( _gvn.transform(result_mem) );
duke@435	2916	push_result(result_reg, result_val);
duke@435	2917	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	2918
duke@435	2919	return true;
duke@435	2920	}
duke@435	2921
duke@435	2922	//----------------------inline_native_getLength--------------------------
duke@435	2923	bool LibraryCallKit::inline_native_getLength() {
duke@435	2924	if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@435	2925
duke@435	2926	int nargs = 1;
duke@435	2927	Node* array = argument(0);
duke@435	2928
duke@435	2929	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2930	array = do_null_check(array, T_OBJECT);
duke@435	2931	_sp -= nargs;
duke@435	2932
duke@435	2933	// If array is dead, only null-path is taken.
duke@435	2934	if (stopped()) return true;
duke@435	2935
duke@435	2936	// Deoptimize if it is a non-array.
duke@435	2937	Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
duke@435	2938
duke@435	2939	if (non_array != NULL) {
duke@435	2940	PreserveJVMState pjvms(this);
duke@435	2941	set_control(non_array);
duke@435	2942	_sp += nargs; // push the arguments back on the stack
duke@435	2943	uncommon_trap(Deoptimization::Reason_intrinsic,
duke@435	2944	Deoptimization::Action_maybe_recompile);
duke@435	2945	}
duke@435	2946
duke@435	2947	// If control is dead, only non-array-path is taken.
duke@435	2948	if (stopped()) return true;
duke@435	2949
duke@435	2950	// The works fine even if the array type is polymorphic.
duke@435	2951	// It could be a dynamic mix of int[], boolean[], Object[], etc.
duke@435	2952	push( load_array_length(array) );
duke@435	2953
duke@435	2954	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	2955
duke@435	2956	return true;
duke@435	2957	}
duke@435	2958
duke@435	2959	//------------------------inline_array_copyOf----------------------------
duke@435	2960	bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
duke@435	2961	if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@435	2962
duke@435	2963	// Restore the stack and pop off the arguments.
duke@435	2964	int nargs = 3 + (is_copyOfRange? 1: 0);
duke@435	2965	Node* original = argument(0);
duke@435	2966	Node* start = is_copyOfRange? argument(1): intcon(0);
duke@435	2967	Node* end = is_copyOfRange? argument(2): argument(1);
duke@435	2968	Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
duke@435	2969
duke@435	2970	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2971	array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
duke@435	2972	original = do_null_check(original, T_OBJECT);
duke@435	2973	_sp -= nargs;
duke@435	2974
duke@435	2975	// Check if a null path was taken unconditionally.
duke@435	2976	if (stopped()) return true;
duke@435	2977
duke@435	2978	Node* orig_length = load_array_length(original);
duke@435	2979
duke@435	2980	Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs,
duke@435	2981	NULL, 0);
duke@435	2982	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2983	klass_node = do_null_check(klass_node, T_OBJECT);
duke@435	2984	_sp -= nargs;
duke@435	2985
duke@435	2986	RegionNode* bailout = new (C, 1) RegionNode(1);
duke@435	2987	record_for_igvn(bailout);
duke@435	2988
duke@435	2989	// Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
duke@435	2990	// Bail out if that is so.
duke@435	2991	Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
duke@435	2992	if (not_objArray != NULL) {
duke@435	2993	// Improve the klass node's type from the new optimistic assumption:
duke@435	2994	ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
duke@435	2995	const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@435	2996	Node* cast = new (C, 2) CastPPNode(klass_node, akls);
duke@435	2997	cast->init_req(0, control());
duke@435	2998	klass_node = _gvn.transform(cast);
duke@435	2999	}
duke@435	3000
duke@435	3001	// Bail out if either start or end is negative.
duke@435	3002	generate_negative_guard(start, bailout, &start);
duke@435	3003	generate_negative_guard(end, bailout, &end);
duke@435	3004
duke@435	3005	Node* length = end;
duke@435	3006	if (_gvn.type(start) != TypeInt::ZERO) {
duke@435	3007	length = _gvn.transform( new (C, 3) SubINode(end, start) );
duke@435	3008	}
duke@435	3009
duke@435	3010	// Bail out if length is negative.
duke@435	3011	// ...Not needed, since the new_array will throw the right exception.
duke@435	3012	//generate_negative_guard(length, bailout, &length);
duke@435	3013
duke@435	3014	if (bailout->req() > 1) {
duke@435	3015	PreserveJVMState pjvms(this);
duke@435	3016	set_control( _gvn.transform(bailout) );
duke@435	3017	_sp += nargs; // push the arguments back on the stack
duke@435	3018	uncommon_trap(Deoptimization::Reason_intrinsic,
duke@435	3019	Deoptimization::Action_maybe_recompile);
duke@435	3020	}
duke@435	3021
duke@435	3022	if (!stopped()) {
duke@435	3023	// How many elements will we copy from the original?
duke@435	3024	// The answer is MinI(orig_length - start, length).
duke@435	3025	Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
duke@435	3026	Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
duke@435	3027
duke@435	3028	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	3029	Node* newcopy = new_array(klass_node, length);
duke@435	3030	_sp -= nargs;
duke@435	3031
duke@435	3032	// Generate a direct call to the right arraycopy function(s).
duke@435	3033	// We know the copy is disjoint but we might not know if the
duke@435	3034	// oop stores need checking.
duke@435	3035	// Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
duke@435	3036	// This will fail a store-check if x contains any non-nulls.
duke@435	3037	bool disjoint_bases = true;
duke@435	3038	bool length_never_negative = true;
duke@435	3039	generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
duke@435	3040	original, start, newcopy, intcon(0), moved,
duke@435	3041	nargs, disjoint_bases, length_never_negative);
duke@435	3042
duke@435	3043	push(newcopy);
duke@435	3044	}
duke@435	3045
duke@435	3046	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	3047
duke@435	3048	return true;
duke@435	3049	}
duke@435	3050
duke@435	3051
duke@435	3052	//----------------------generate_virtual_guard---------------------------
duke@435	3053	// Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
duke@435	3054	Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
duke@435	3055	RegionNode* slow_region) {
duke@435	3056	ciMethod* method = callee();
duke@435	3057	int vtable_index = method->vtable_index();
duke@435	3058	// Get the methodOop out of the appropriate vtable entry.
duke@435	3059	int entry_offset = (instanceKlass::vtable_start_offset() +
duke@435	3060	vtable_index*vtableEntry::size()) * wordSize +
duke@435	3061	vtableEntry::method_offset_in_bytes();
duke@435	3062	Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
duke@435	3063	Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
duke@435	3064
duke@435	3065	// Compare the target method with the expected method (e.g., Object.hashCode).
duke@435	3066	const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
duke@435	3067
duke@435	3068	Node* native_call = makecon(native_call_addr);
duke@435	3069	Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
duke@435	3070	Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
duke@435	3071
duke@435	3072	return generate_slow_guard(test_native, slow_region);
duke@435	3073	}
duke@435	3074
duke@435	3075	//-----------------------generate_method_call----------------------------
duke@435	3076	// Use generate_method_call to make a slow-call to the real
duke@435	3077	// method if the fast path fails. An alternative would be to
duke@435	3078	// use a stub like OptoRuntime::slow_arraycopy_Java.
duke@435	3079	// This only works for expanding the current library call,
duke@435	3080	// not another intrinsic. (E.g., don't use this for making an
duke@435	3081	// arraycopy call inside of the copyOf intrinsic.)
duke@435	3082	CallJavaNode*
duke@435	3083	LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
duke@435	3084	// When compiling the intrinsic method itself, do not use this technique.
duke@435	3085	guarantee(callee() != C->method(), "cannot make slow-call to self");
duke@435	3086
duke@435	3087	ciMethod* method = callee();
duke@435	3088	// ensure the JVMS we have will be correct for this call
duke@435	3089	guarantee(method_id == method->intrinsic_id(), "must match");
duke@435	3090
duke@435	3091	const TypeFunc* tf = TypeFunc::make(method);
duke@435	3092	int tfdc = tf->domain()->cnt();
duke@435	3093	CallJavaNode* slow_call;
duke@435	3094	if (is_static) {
duke@435	3095	assert(!is_virtual, "");
duke@435	3096	slow_call = new(C, tfdc) CallStaticJavaNode(tf,
duke@435	3097	SharedRuntime::get_resolve_static_call_stub(),
duke@435	3098	method, bci());
duke@435	3099	} else if (is_virtual) {
duke@435	3100	null_check_receiver(method);
duke@435	3101	int vtable_index = methodOopDesc::invalid_vtable_index;
duke@435	3102	if (UseInlineCaches) {
duke@435	3103	// Suppress the vtable call
duke@435	3104	} else {
duke@435	3105	// hashCode and clone are not a miranda methods,
duke@435	3106	// so the vtable index is fixed.
duke@435	3107	// No need to use the linkResolver to get it.
duke@435	3108	vtable_index = method->vtable_index();
duke@435	3109	}
duke@435	3110	slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
duke@435	3111	SharedRuntime::get_resolve_virtual_call_stub(),
duke@435	3112	method, vtable_index, bci());
duke@435	3113	} else { // neither virtual nor static: opt_virtual
duke@435	3114	null_check_receiver(method);
duke@435	3115	slow_call = new(C, tfdc) CallStaticJavaNode(tf,
duke@435	3116	SharedRuntime::get_resolve_opt_virtual_call_stub(),
duke@435	3117	method, bci());
duke@435	3118	slow_call->set_optimized_virtual(true);
duke@435	3119	}
duke@435	3120	set_arguments_for_java_call(slow_call);
duke@435	3121	set_edges_for_java_call(slow_call);
duke@435	3122	return slow_call;
duke@435	3123	}
duke@435	3124
duke@435	3125
duke@435	3126	//------------------------------inline_native_hashcode--------------------
duke@435	3127	// Build special case code for calls to hashCode on an object.
duke@435	3128	bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
duke@435	3129	assert(is_static == callee()->is_static(), "correct intrinsic selection");
duke@435	3130	assert(!(is_virtual && is_static), "either virtual, special, or static");
duke@435	3131
duke@435	3132	enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
duke@435	3133
duke@435	3134	RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	3135	PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
duke@435	3136	TypeInt::INT);
duke@435	3137	PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
duke@435	3138	PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
duke@435	3139	TypePtr::BOTTOM);
duke@435	3140	Node* obj = NULL;
duke@435	3141	if (!is_static) {
duke@435	3142	// Check for hashing null object
duke@435	3143	obj = null_check_receiver(callee());
duke@435	3144	if (stopped()) return true; // unconditionally null
duke@435	3145	result_reg->init_req(_null_path, top());
duke@435	3146	result_val->init_req(_null_path, top());
duke@435	3147	} else {
duke@435	3148	// Do a null check, and return zero if null.
duke@435	3149	// System.identityHashCode(null) == 0
duke@435	3150	obj = argument(0);
duke@435	3151	Node* null_ctl = top();
duke@435	3152	obj = null_check_oop(obj, &null_ctl);
duke@435	3153	result_reg->init_req(_null_path, null_ctl);
duke@435	3154	result_val->init_req(_null_path, _gvn.intcon(0));
duke@435	3155	}
duke@435	3156
duke@435	3157	// Unconditionally null? Then return right away.
duke@435	3158	if (stopped()) {
duke@435	3159	set_control( result_reg->in(_null_path) );
duke@435	3160	if (!stopped())
duke@435	3161	push( result_val ->in(_null_path) );
duke@435	3162	return true;
duke@435	3163	}
duke@435	3164
duke@435	3165	// After null check, get the object's klass.
duke@435	3166	Node* obj_klass = load_object_klass(obj);
duke@435	3167
duke@435	3168	// This call may be virtual (invokevirtual) or bound (invokespecial).
duke@435	3169	// For each case we generate slightly different code.
duke@435	3170
duke@435	3171	// We only go to the fast case code if we pass a number of guards. The
duke@435	3172	// paths which do not pass are accumulated in the slow_region.
duke@435	3173	RegionNode* slow_region = new (C, 1) RegionNode(1);
duke@435	3174	record_for_igvn(slow_region);
duke@435	3175
duke@435	3176	// If this is a virtual call, we generate a funny guard. We pull out
duke@435	3177	// the vtable entry corresponding to hashCode() from the target object.
duke@435	3178	// If the target method which we are calling happens to be the native
duke@435	3179	// Object hashCode() method, we pass the guard. We do not need this
duke@435	3180	// guard for non-virtual calls -- the caller is known to be the native
duke@435	3181	// Object hashCode().
duke@435	3182	if (is_virtual) {
duke@435	3183	generate_virtual_guard(obj_klass, slow_region);
duke@435	3184	}
duke@435	3185
duke@435	3186	// Get the header out of the object, use LoadMarkNode when available
duke@435	3187	Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
duke@435	3188	Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
duke@435	3189	header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
duke@435	3190
duke@435	3191	// Test the header to see if it is unlocked.
duke@435	3192	Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
duke@435	3193	Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
duke@435	3194	Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
duke@435	3195	Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
duke@435	3196	Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
duke@435	3197
duke@435	3198	generate_slow_guard(test_unlocked, slow_region);
duke@435	3199
duke@435	3200	// Get the hash value and check to see that it has been properly assigned.
duke@435	3201	// We depend on hash_mask being at most 32 bits and avoid the use of
duke@435	3202	// hash_mask_in_place because it could be larger than 32 bits in a 64-bit
duke@435	3203	// vm: see markOop.hpp.
duke@435	3204	Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
duke@435	3205	Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
duke@435	3206	Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
duke@435	3207	// This hack lets the hash bits live anywhere in the mark object now, as long
duke@435	3208	// as the shift drops the relevent bits into the low 32 bits. Note that
duke@435	3209	// Java spec says that HashCode is an int so there's no point in capturing
duke@435	3210	// an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
duke@435	3211	hshifted_header = ConvX2I(hshifted_header);
duke@435	3212	Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
duke@435	3213
duke@435	3214	Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
duke@435	3215	Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
duke@435	3216	Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
duke@435	3217
duke@435	3218	generate_slow_guard(test_assigned, slow_region);
duke@435	3219
duke@435	3220	Node* init_mem = reset_memory();
duke@435	3221	// fill in the rest of the null path:
duke@435	3222	result_io ->init_req(_null_path, i_o());
duke@435	3223	result_mem->init_req(_null_path, init_mem);
duke@435	3224
duke@435	3225	result_val->init_req(_fast_path, hash_val);
duke@435	3226	result_reg->init_req(_fast_path, control());
duke@435	3227	result_io ->init_req(_fast_path, i_o());
duke@435	3228	result_mem->init_req(_fast_path, init_mem);
duke@435	3229
duke@435	3230	// Generate code for the slow case. We make a call to hashCode().
duke@435	3231	set_control(_gvn.transform(slow_region));
duke@435	3232	if (!stopped()) {
duke@435	3233	// No need for PreserveJVMState, because we're using up the present state.
duke@435	3234	set_all_memory(init_mem);
duke@435	3235	vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
duke@435	3236	if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
duke@435	3237	CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
duke@435	3238	Node* slow_result = set_results_for_java_call(slow_call);
duke@435	3239	// this->control() comes from set_results_for_java_call
duke@435	3240	result_reg->init_req(_slow_path, control());
duke@435	3241	result_val->init_req(_slow_path, slow_result);
duke@435	3242	result_io ->set_req(_slow_path, i_o());
duke@435	3243	result_mem ->set_req(_slow_path, reset_memory());
duke@435	3244	}
duke@435	3245
duke@435	3246	// Return the combined state.
duke@435	3247	set_i_o( _gvn.transform(result_io) );
duke@435	3248	set_all_memory( _gvn.transform(result_mem) );
duke@435	3249	push_result(result_reg, result_val);
duke@435	3250
duke@435	3251	return true;
duke@435	3252	}
duke@435	3253
duke@435	3254	//---------------------------inline_native_getClass----------------------------
duke@435	3255	// Build special case code for calls to hashCode on an object.
duke@435	3256	bool LibraryCallKit::inline_native_getClass() {
duke@435	3257	Node* obj = null_check_receiver(callee());
duke@435	3258	if (stopped()) return true;
duke@435	3259	push( load_mirror_from_klass(load_object_klass(obj)) );
duke@435	3260	return true;
duke@435	3261	}
duke@435	3262
duke@435	3263	//-----------------inline_native_Reflection_getCallerClass---------------------
duke@435	3264	// In the presence of deep enough inlining, getCallerClass() becomes a no-op.
duke@435	3265	//
duke@435	3266	// NOTE that this code must perform the same logic as
duke@435	3267	// vframeStream::security_get_caller_frame in that it must skip
duke@435	3268	// Method.invoke() and auxiliary frames.
duke@435	3269
duke@435	3270
duke@435	3271
duke@435	3272
duke@435	3273	bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
duke@435	3274	ciMethod* method = callee();
duke@435	3275
duke@435	3276	#ifndef PRODUCT
duke@435	3277	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3278	tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
duke@435	3279	}
duke@435	3280	#endif
duke@435	3281
duke@435	3282	debug_only(int saved_sp = _sp);
duke@435	3283
duke@435	3284	// Argument words: (int depth)
duke@435	3285	int nargs = 1;
duke@435	3286
duke@435	3287	_sp += nargs;
duke@435	3288	Node* caller_depth_node = pop();
duke@435	3289
duke@435	3290	assert(saved_sp == _sp, "must have correct argument count");
duke@435	3291
duke@435	3292	// The depth value must be a constant in order for the runtime call
duke@435	3293	// to be eliminated.
duke@435	3294	const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
duke@435	3295	if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
duke@435	3296	#ifndef PRODUCT
duke@435	3297	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3298	tty->print_cr(" Bailing out because caller depth was not a constant");
duke@435	3299	}
duke@435	3300	#endif
duke@435	3301	return false;
duke@435	3302	}
duke@435	3303	// Note that the JVM state at this point does not include the
duke@435	3304	// getCallerClass() frame which we are trying to inline. The
duke@435	3305	// semantics of getCallerClass(), however, are that the "first"
duke@435	3306	// frame is the getCallerClass() frame, so we subtract one from the
duke@435	3307	// requested depth before continuing. We don't inline requests of
duke@435	3308	// getCallerClass(0).
duke@435	3309	int caller_depth = caller_depth_type->get_con() - 1;
duke@435	3310	if (caller_depth < 0) {
duke@435	3311	#ifndef PRODUCT
duke@435	3312	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3313	tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
duke@435	3314	}
duke@435	3315	#endif
duke@435	3316	return false;
duke@435	3317	}
duke@435	3318
duke@435	3319	if (!jvms()->has_method()) {
duke@435	3320	#ifndef PRODUCT
duke@435	3321	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3322	tty->print_cr(" Bailing out because intrinsic was inlined at top level");
duke@435	3323	}
duke@435	3324	#endif
duke@435	3325	return false;
duke@435	3326	}
duke@435	3327	int _depth = jvms()->depth(); // cache call chain depth
duke@435	3328
duke@435	3329	// Walk back up the JVM state to find the caller at the required
duke@435	3330	// depth. NOTE that this code must perform the same logic as
duke@435	3331	// vframeStream::security_get_caller_frame in that it must skip
duke@435	3332	// Method.invoke() and auxiliary frames. Note also that depth is
duke@435	3333	// 1-based (1 is the bottom of the inlining).
duke@435	3334	int inlining_depth = _depth;
duke@435	3335	JVMState* caller_jvms = NULL;
duke@435	3336
duke@435	3337	if (inlining_depth > 0) {
duke@435	3338	caller_jvms = jvms();
duke@435	3339	assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
duke@435	3340	do {
duke@435	3341	// The following if-tests should be performed in this order
duke@435	3342	if (is_method_invoke_or_aux_frame(caller_jvms)) {
duke@435	3343	// Skip a Method.invoke() or auxiliary frame
duke@435	3344	} else if (caller_depth > 0) {
duke@435	3345	// Skip real frame
duke@435	3346	--caller_depth;
duke@435	3347	} else {
duke@435	3348	// We're done: reached desired caller after skipping.
duke@435	3349	break;
duke@435	3350	}
duke@435	3351	caller_jvms = caller_jvms->caller();
duke@435	3352	--inlining_depth;
duke@435	3353	} while (inlining_depth > 0);
duke@435	3354	}
duke@435	3355
duke@435	3356	if (inlining_depth == 0) {
duke@435	3357	#ifndef PRODUCT
duke@435	3358	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3359	tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
duke@435	3360	tty->print_cr(" JVM state at this point:");
duke@435	3361	for (int i = _depth; i >= 1; i--) {
duke@435	3362	tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
duke@435	3363	}
duke@435	3364	}
duke@435	3365	#endif
duke@435	3366	return false; // Reached end of inlining
duke@435	3367	}
duke@435	3368
duke@435	3369	// Acquire method holder as java.lang.Class
duke@435	3370	ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
duke@435	3371	ciInstance* caller_mirror = caller_klass->java_mirror();
duke@435	3372	// Push this as a constant
duke@435	3373	push(makecon(TypeInstPtr::make(caller_mirror)));
duke@435	3374	#ifndef PRODUCT
duke@435	3375	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3376	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	3377	tty->print_cr(" JVM state at this point:");
duke@435	3378	for (int i = _depth; i >= 1; i--) {
duke@435	3379	tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
duke@435	3380	}
duke@435	3381	}
duke@435	3382	#endif
duke@435	3383	return true;
duke@435	3384	}
duke@435	3385
duke@435	3386	// Helper routine for above
duke@435	3387	bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
duke@435	3388	// Is this the Method.invoke method itself?
duke@435	3389	if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke)
duke@435	3390	return true;
duke@435	3391
duke@435	3392	// Is this a helper, defined somewhere underneath MethodAccessorImpl.
duke@435	3393	ciKlass* k = jvms->method()->holder();
duke@435	3394	if (k->is_instance_klass()) {
duke@435	3395	ciInstanceKlass* ik = k->as_instance_klass();
duke@435	3396	for (; ik != NULL; ik = ik->super()) {
duke@435	3397	if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
duke@435	3398	ik == env()->find_system_klass(ik->name())) {
duke@435	3399	return true;
duke@435	3400	}
duke@435	3401	}
duke@435	3402	}
duke@435	3403
duke@435	3404	return false;
duke@435	3405	}
duke@435	3406
duke@435	3407	static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
duke@435	3408	// inline_native_AtomicLong_attemptUpdate() but it has no way of
duke@435	3409	// computing it since there is no lookup field by name function in the
duke@435	3410	// CI interface. This is computed and set by inline_native_AtomicLong_get().
duke@435	3411	// Using a static variable here is safe even if we have multiple compilation
duke@435	3412	// threads because the offset is constant. At worst the same offset will be
duke@435	3413	// computed and stored multiple
duke@435	3414
duke@435	3415	bool LibraryCallKit::inline_native_AtomicLong_get() {
duke@435	3416	// Restore the stack and pop off the argument
duke@435	3417	_sp+=1;
duke@435	3418	Node *obj = pop();
duke@435	3419
duke@435	3420	// get the offset of the "value" field. Since the CI interfaces
duke@435	3421	// does not provide a way to look up a field by name, we scan the bytecodes
duke@435	3422	// to get the field index. We expect the first 2 instructions of the method
duke@435	3423	// to be:
duke@435	3424	// 0 aload_0
duke@435	3425	// 1 getfield "value"
duke@435	3426	ciMethod* method = callee();
duke@435	3427	if (value_field_offset == -1)
duke@435	3428	{
duke@435	3429	ciField* value_field;
duke@435	3430	ciBytecodeStream iter(method);
duke@435	3431	Bytecodes::Code bc = iter.next();
duke@435	3432
duke@435	3433	if ((bc != Bytecodes::_aload_0) &&
duke@435	3434	((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
duke@435	3435	return false;
duke@435	3436	bc = iter.next();
duke@435	3437	if (bc != Bytecodes::_getfield)
duke@435	3438	return false;
duke@435	3439	bool ignore;
duke@435	3440	value_field = iter.get_field(ignore);
duke@435	3441	value_field_offset = value_field->offset_in_bytes();
duke@435	3442	}
duke@435	3443
duke@435	3444	// Null check without removing any arguments.
duke@435	3445	_sp++;
duke@435	3446	obj = do_null_check(obj, T_OBJECT);
duke@435	3447	_sp--;
duke@435	3448	// Check for locking null object
duke@435	3449	if (stopped()) return true;
duke@435	3450
duke@435	3451	Node *adr = basic_plus_adr(obj, obj, value_field_offset);
duke@435	3452	const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
duke@435	3453	int alias_idx = C->get_alias_index(adr_type);
duke@435	3454
duke@435	3455	Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
duke@435	3456
duke@435	3457	push_pair(result);
duke@435	3458
duke@435	3459	return true;
duke@435	3460	}
duke@435	3461
duke@435	3462	bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
duke@435	3463	// Restore the stack and pop off the arguments
duke@435	3464	_sp+=5;
duke@435	3465	Node *newVal = pop_pair();
duke@435	3466	Node *oldVal = pop_pair();
duke@435	3467	Node *obj = pop();
duke@435	3468
duke@435	3469	// we need the offset of the "value" field which was computed when
duke@435	3470	// inlining the get() method. Give up if we don't have it.
duke@435	3471	if (value_field_offset == -1)
duke@435	3472	return false;
duke@435	3473
duke@435	3474	// Null check without removing any arguments.
duke@435	3475	_sp+=5;
duke@435	3476	obj = do_null_check(obj, T_OBJECT);
duke@435	3477	_sp-=5;
duke@435	3478	// Check for locking null object
duke@435	3479	if (stopped()) return true;
duke@435	3480
duke@435	3481	Node *adr = basic_plus_adr(obj, obj, value_field_offset);
duke@435	3482	const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
duke@435	3483	int alias_idx = C->get_alias_index(adr_type);
duke@435	3484
duke@435	3485	Node *result = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
duke@435	3486	Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(result));
duke@435	3487	set_memory(store_proj, alias_idx);
duke@435	3488
duke@435	3489	push(result);
duke@435	3490	return true;
duke@435	3491	}
duke@435	3492
duke@435	3493	bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
duke@435	3494	// restore the arguments
duke@435	3495	_sp += arg_size();
duke@435	3496
duke@435	3497	switch (id) {
duke@435	3498	case vmIntrinsics::_floatToRawIntBits:
duke@435	3499	push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
duke@435	3500	break;
duke@435	3501
duke@435	3502	case vmIntrinsics::_intBitsToFloat:
duke@435	3503	push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
duke@435	3504	break;
duke@435	3505
duke@435	3506	case vmIntrinsics::_doubleToRawLongBits:
duke@435	3507	push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
duke@435	3508	break;
duke@435	3509
duke@435	3510	case vmIntrinsics::_longBitsToDouble:
duke@435	3511	push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
duke@435	3512	break;
duke@435	3513
duke@435	3514	case vmIntrinsics::_doubleToLongBits: {
duke@435	3515	Node* value = pop_pair();
duke@435	3516
duke@435	3517	// two paths (plus control) merge in a wood
duke@435	3518	RegionNode *r = new (C, 3) RegionNode(3);
duke@435	3519	Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
duke@435	3520
duke@435	3521	Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
duke@435	3522	// Build the boolean node
duke@435	3523	Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
duke@435	3524
duke@435	3525	// Branch either way.
duke@435	3526	// NaN case is less traveled, which makes all the difference.
duke@435	3527	IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
duke@435	3528	Node *opt_isnan = _gvn.transform(ifisnan);
duke@435	3529	assert( opt_isnan->is_If(), "Expect an IfNode");
duke@435	3530	IfNode *opt_ifisnan = (IfNode*)opt_isnan;
duke@435	3531	Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
duke@435	3532
duke@435	3533	set_control(iftrue);
duke@435	3534
duke@435	3535	static const jlong nan_bits = CONST64(0x7ff8000000000000);
duke@435	3536	Node *slow_result = longcon(nan_bits); // return NaN
duke@435	3537	phi->init_req(1, _gvn.transform( slow_result ));
duke@435	3538	r->init_req(1, iftrue);
duke@435	3539
duke@435	3540	// Else fall through
duke@435	3541	Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
duke@435	3542	set_control(iffalse);
duke@435	3543
duke@435	3544	phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
duke@435	3545	r->init_req(2, iffalse);
duke@435	3546
duke@435	3547	// Post merge
duke@435	3548	set_control(_gvn.transform(r));
duke@435	3549	record_for_igvn(r);
duke@435	3550
duke@435	3551	Node* result = _gvn.transform(phi);
duke@435	3552	assert(result->bottom_type()->isa_long(), "must be");
duke@435	3553	push_pair(result);
duke@435	3554
duke@435	3555	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	3556
duke@435	3557	break;
duke@435	3558	}
duke@435	3559
duke@435	3560	case vmIntrinsics::_floatToIntBits: {
duke@435	3561	Node* value = pop();
duke@435	3562
duke@435	3563	// two paths (plus control) merge in a wood
duke@435	3564	RegionNode *r = new (C, 3) RegionNode(3);
duke@435	3565	Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
duke@435	3566
duke@435	3567	Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
duke@435	3568	// Build the boolean node
duke@435	3569	Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
duke@435	3570
duke@435	3571	// Branch either way.
duke@435	3572	// NaN case is less traveled, which makes all the difference.
duke@435	3573	IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
duke@435	3574	Node *opt_isnan = _gvn.transform(ifisnan);
duke@435	3575	assert( opt_isnan->is_If(), "Expect an IfNode");
duke@435	3576	IfNode *opt_ifisnan = (IfNode*)opt_isnan;
duke@435	3577	Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
duke@435	3578
duke@435	3579	set_control(iftrue);
duke@435	3580
duke@435	3581	static const jint nan_bits = 0x7fc00000;
duke@435	3582	Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
duke@435	3583	phi->init_req(1, _gvn.transform( slow_result ));
duke@435	3584	r->init_req(1, iftrue);
duke@435	3585
duke@435	3586	// Else fall through
duke@435	3587	Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
duke@435	3588	set_control(iffalse);
duke@435	3589
duke@435	3590	phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
duke@435	3591	r->init_req(2, iffalse);
duke@435	3592
duke@435	3593	// Post merge
duke@435	3594	set_control(_gvn.transform(r));
duke@435	3595	record_for_igvn(r);
duke@435	3596
duke@435	3597	Node* result = _gvn.transform(phi);
duke@435	3598	assert(result->bottom_type()->isa_int(), "must be");
duke@435	3599	push(result);
duke@435	3600
duke@435	3601	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	3602
duke@435	3603	break;
duke@435	3604	}
duke@435	3605
duke@435	3606	default:
duke@435	3607	ShouldNotReachHere();
duke@435	3608	}
duke@435	3609
duke@435	3610	return true;
duke@435	3611	}
duke@435	3612
duke@435	3613	#ifdef _LP64
duke@435	3614	#define XTOP ,top() /*additional argument*/
duke@435	3615	#else //_LP64
duke@435	3616	#define XTOP /*no additional argument*/
duke@435	3617	#endif //_LP64
duke@435	3618
duke@435	3619	//----------------------inline_unsafe_copyMemory-------------------------
duke@435	3620	bool LibraryCallKit::inline_unsafe_copyMemory() {
duke@435	3621	if (callee()->is_static()) return false; // caller must have the capability!
duke@435	3622	int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
duke@435	3623	assert(signature()->size() == nargs-1, "copy has 5 arguments");
duke@435	3624	null_check_receiver(callee()); // check then ignore argument(0)
duke@435	3625	if (stopped()) return true;
duke@435	3626
duke@435	3627	C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@435	3628
duke@435	3629	Node* src_ptr = argument(1);
duke@435	3630	Node* src_off = ConvL2X(argument(2));
duke@435	3631	assert(argument(3)->is_top(), "2nd half of long");
duke@435	3632	Node* dst_ptr = argument(4);
duke@435	3633	Node* dst_off = ConvL2X(argument(5));
duke@435	3634	assert(argument(6)->is_top(), "2nd half of long");
duke@435	3635	Node* size = ConvL2X(argument(7));
duke@435	3636	assert(argument(8)->is_top(), "2nd half of long");
duke@435	3637
duke@435	3638	assert(Unsafe_field_offset_to_byte_offset(11) == 11,
duke@435	3639	"fieldOffset must be byte-scaled");
duke@435	3640
duke@435	3641	Node* src = make_unsafe_address(src_ptr, src_off);
duke@435	3642	Node* dst = make_unsafe_address(dst_ptr, dst_off);
duke@435	3643
duke@435	3644	// Conservatively insert a memory barrier on all memory slices.
duke@435	3645	// Do not let writes of the copy source or destination float below the copy.
duke@435	3646	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	3647
duke@435	3648	// Call it. Note that the length argument is not scaled.
duke@435	3649	make_runtime_call(RC_LEAF|RC_NO_FP,
duke@435	3650	OptoRuntime::fast_arraycopy_Type(),
duke@435	3651	StubRoutines::unsafe_arraycopy(),
duke@435	3652	"unsafe_arraycopy",
duke@435	3653	TypeRawPtr::BOTTOM,
duke@435	3654	src, dst, size XTOP);
duke@435	3655
duke@435	3656	// Do not let reads of the copy destination float above the copy.
duke@435	3657	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	3658
duke@435	3659	return true;
duke@435	3660	}
duke@435	3661
duke@435	3662
duke@435	3663	//------------------------inline_native_clone----------------------------
duke@435	3664	// Here are the simple edge cases:
duke@435	3665	// null receiver => normal trap
duke@435	3666	// virtual and clone was overridden => slow path to out-of-line clone
duke@435	3667	// not cloneable or finalizer => slow path to out-of-line Object.clone
duke@435	3668	//
duke@435	3669	// The general case has two steps, allocation and copying.
duke@435	3670	// Allocation has two cases, and uses GraphKit::new_instance or new_array.
duke@435	3671	//
duke@435	3672	// Copying also has two cases, oop arrays and everything else.
duke@435	3673	// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
duke@435	3674	// Everything else uses the tight inline loop supplied by CopyArrayNode.
duke@435	3675	//
duke@435	3676	// These steps fold up nicely if and when the cloned object's klass
duke@435	3677	// can be sharply typed as an object array, a type array, or an instance.
duke@435	3678	//
duke@435	3679	bool LibraryCallKit::inline_native_clone(bool is_virtual) {
duke@435	3680	int nargs = 1;
duke@435	3681	Node* obj = null_check_receiver(callee());
duke@435	3682	if (stopped()) return true;
duke@435	3683	Node* obj_klass = load_object_klass(obj);
duke@435	3684	const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
duke@435	3685	const TypeOopPtr* toop = ((tklass != NULL)
duke@435	3686	? tklass->as_instance_type()
duke@435	3687	: TypeInstPtr::NOTNULL);
duke@435	3688
duke@435	3689	// Conservatively insert a memory barrier on all memory slices.
duke@435	3690	// Do not let writes into the original float below the clone.
duke@435	3691	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	3692
duke@435	3693	// paths into result_reg:
duke@435	3694	enum {
duke@435	3695	_slow_path = 1, // out-of-line call to clone method (virtual or not)
duke@435	3696	_objArray_path, // plain allocation, plus arrayof_oop_arraycopy
duke@435	3697	_fast_path, // plain allocation, plus a CopyArray operation
duke@435	3698	PATH_LIMIT
duke@435	3699	};
duke@435	3700	RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	3701	PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
duke@435	3702	TypeInstPtr::NOTNULL);
duke@435	3703	PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
duke@435	3704	PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
duke@435	3705	TypePtr::BOTTOM);
duke@435	3706	record_for_igvn(result_reg);
duke@435	3707
duke@435	3708	const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
duke@435	3709	int raw_adr_idx = Compile::AliasIdxRaw;
duke@435	3710	const bool raw_mem_only = true;
duke@435	3711
duke@435	3712	// paths into alloc_reg (on the fast path, just before the CopyArray):
duke@435	3713	enum { _typeArray_alloc = 1, _instance_alloc, ALLOC_LIMIT };
duke@435	3714	RegionNode* alloc_reg = new(C, ALLOC_LIMIT) RegionNode(ALLOC_LIMIT);
duke@435	3715	PhiNode* alloc_val = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, raw_adr_type);
duke@435	3716	PhiNode* alloc_siz = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, TypeX_X);
duke@435	3717	PhiNode* alloc_i_o = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::ABIO);
duke@435	3718	PhiNode* alloc_mem = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::MEMORY,
duke@435	3719	raw_adr_type);
duke@435	3720	record_for_igvn(alloc_reg);
duke@435	3721
duke@435	3722	bool card_mark = false; // (see below)
duke@435	3723
duke@435	3724	Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
duke@435	3725	if (array_ctl != NULL) {
duke@435	3726	// It's an array.
duke@435	3727	PreserveJVMState pjvms(this);
duke@435	3728	set_control(array_ctl);
duke@435	3729	Node* obj_length = load_array_length(obj);
duke@435	3730	Node* obj_size = NULL;
duke@435	3731	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	3732	Node* alloc_obj = new_array(obj_klass, obj_length,
duke@435	3733	raw_mem_only, &obj_size);
duke@435	3734	_sp -= nargs;
duke@435	3735	assert(obj_size != NULL, "");
duke@435	3736	Node* raw_obj = alloc_obj->in(1);
duke@435	3737	assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
duke@435	3738	if (ReduceBulkZeroing) {
duke@435	3739	AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
duke@435	3740	if (alloc != NULL) {
duke@435	3741	// We will be completely responsible for initializing this object.
duke@435	3742	alloc->maybe_set_complete(&_gvn);
duke@435	3743	}
duke@435	3744	}
duke@435	3745
duke@435	3746	if (!use_ReduceInitialCardMarks()) {
duke@435	3747	// If it is an oop array, it requires very special treatment,
duke@435	3748	// because card marking is required on each card of the array.
duke@435	3749	Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
duke@435	3750	if (is_obja != NULL) {
duke@435	3751	PreserveJVMState pjvms2(this);
duke@435	3752	set_control(is_obja);
duke@435	3753	// Generate a direct call to the right arraycopy function(s).
duke@435	3754	bool disjoint_bases = true;
duke@435	3755	bool length_never_negative = true;
duke@435	3756	generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
duke@435	3757	obj, intcon(0), alloc_obj, intcon(0),
duke@435	3758	obj_length, nargs,
duke@435	3759	disjoint_bases, length_never_negative);
duke@435	3760	result_reg->init_req(_objArray_path, control());
duke@435	3761	result_val->init_req(_objArray_path, alloc_obj);
duke@435	3762	result_i_o ->set_req(_objArray_path, i_o());
duke@435	3763	result_mem ->set_req(_objArray_path, reset_memory());
duke@435	3764	}
duke@435	3765	}
duke@435	3766	// We can dispense with card marks if we know the allocation
duke@435	3767	// comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
duke@435	3768	// causes the non-eden paths to simulate a fresh allocation,
duke@435	3769	// insofar that no further card marks are required to initialize
duke@435	3770	// the object.
duke@435	3771
duke@435	3772	// Otherwise, there are no card marks to worry about.
duke@435	3773	alloc_val->init_req(_typeArray_alloc, raw_obj);
duke@435	3774	alloc_siz->init_req(_typeArray_alloc, obj_size);
duke@435	3775	alloc_reg->init_req(_typeArray_alloc, control());
duke@435	3776	alloc_i_o->init_req(_typeArray_alloc, i_o());
duke@435	3777	alloc_mem->init_req(_typeArray_alloc, memory(raw_adr_type));
duke@435	3778	}
duke@435	3779
duke@435	3780	// We only go to the fast case code if we pass a number of guards.
duke@435	3781	// The paths which do not pass are accumulated in the slow_region.
duke@435	3782	RegionNode* slow_region = new (C, 1) RegionNode(1);
duke@435	3783	record_for_igvn(slow_region);
duke@435	3784	if (!stopped()) {
duke@435	3785	// It's an instance. Make the slow-path tests.
duke@435	3786	// If this is a virtual call, we generate a funny guard. We grab
duke@435	3787	// the vtable entry corresponding to clone() from the target object.
duke@435	3788	// If the target method which we are calling happens to be the
duke@435	3789	// Object clone() method, we pass the guard. We do not need this
duke@435	3790	// guard for non-virtual calls; the caller is known to be the native
duke@435	3791	// Object clone().
duke@435	3792	if (is_virtual) {
duke@435	3793	generate_virtual_guard(obj_klass, slow_region);
duke@435	3794	}
duke@435	3795
duke@435	3796	// The object must be cloneable and must not have a finalizer.
duke@435	3797	// Both of these conditions may be checked in a single test.
duke@435	3798	// We could optimize the cloneable test further, but we don't care.
duke@435	3799	generate_access_flags_guard(obj_klass,
duke@435	3800	// Test both conditions:
duke@435	3801	JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
duke@435	3802	// Must be cloneable but not finalizer:
duke@435	3803	JVM_ACC_IS_CLONEABLE,
duke@435	3804	slow_region);
duke@435	3805	}
duke@435	3806
duke@435	3807	if (!stopped()) {
duke@435	3808	// It's an instance, and it passed the slow-path tests.
duke@435	3809	PreserveJVMState pjvms(this);
duke@435	3810	Node* obj_size = NULL;
duke@435	3811	Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
duke@435	3812	assert(obj_size != NULL, "");
duke@435	3813	Node* raw_obj = alloc_obj->in(1);
duke@435	3814	assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
duke@435	3815	if (ReduceBulkZeroing) {
duke@435	3816	AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
duke@435	3817	if (alloc != NULL && !alloc->maybe_set_complete(&_gvn))
duke@435	3818	alloc = NULL;
duke@435	3819	}
duke@435	3820	if (!use_ReduceInitialCardMarks()) {
duke@435	3821	// Put in store barrier for any and all oops we are sticking
duke@435	3822	// into this object. (We could avoid this if we could prove
duke@435	3823	// that the object type contains no oop fields at all.)
duke@435	3824	card_mark = true;
duke@435	3825	}
duke@435	3826	alloc_val->init_req(_instance_alloc, raw_obj);
duke@435	3827	alloc_siz->init_req(_instance_alloc, obj_size);
duke@435	3828	alloc_reg->init_req(_instance_alloc, control());
duke@435	3829	alloc_i_o->init_req(_instance_alloc, i_o());
duke@435	3830	alloc_mem->init_req(_instance_alloc, memory(raw_adr_type));
duke@435	3831	}
duke@435	3832
duke@435	3833	// Generate code for the slow case. We make a call to clone().
duke@435	3834	set_control(_gvn.transform(slow_region));
duke@435	3835	if (!stopped()) {
duke@435	3836	PreserveJVMState pjvms(this);
duke@435	3837	CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
duke@435	3838	Node* slow_result = set_results_for_java_call(slow_call);
duke@435	3839	// this->control() comes from set_results_for_java_call
duke@435	3840	result_reg->init_req(_slow_path, control());
duke@435	3841	result_val->init_req(_slow_path, slow_result);
duke@435	3842	result_i_o ->set_req(_slow_path, i_o());
duke@435	3843	result_mem ->set_req(_slow_path, reset_memory());
duke@435	3844	}
duke@435	3845
duke@435	3846	// The object is allocated, as an array and/or an instance. Now copy it.
duke@435	3847	set_control( _gvn.transform(alloc_reg) );
duke@435	3848	set_i_o( _gvn.transform(alloc_i_o) );
duke@435	3849	set_memory( _gvn.transform(alloc_mem), raw_adr_type );
duke@435	3850	Node* raw_obj = _gvn.transform(alloc_val);
duke@435	3851
duke@435	3852	if (!stopped()) {
duke@435	3853	// Copy the fastest available way.
duke@435	3854	// (No need for PreserveJVMState, since we're using it all up now.)
kvn@598	3855	// TODO: generate fields/elements copies for small objects instead.
duke@435	3856	Node* src = obj;
duke@435	3857	Node* dest = raw_obj;
duke@435	3858	Node* size = _gvn.transform(alloc_siz);
duke@435	3859
duke@435	3860	// Exclude the header.
coleenp@548	3861	int base_off = instanceOopDesc::base_offset_in_bytes();
coleenp@548	3862	if (UseCompressedOops) {
kvn@598	3863	assert(base_off % BytesPerLong != 0, "base with compressed oops");
kvn@598	3864	// With compressed oops base_offset_in_bytes is 12 which creates
kvn@598	3865	// the gap since countx is rounded by 8 bytes below.
kvn@598	3866	// Copy klass and the gap.
kvn@598	3867	base_off = instanceOopDesc::klass_offset_in_bytes();
coleenp@548	3868	}
duke@435	3869	src = basic_plus_adr(src, base_off);
duke@435	3870	dest = basic_plus_adr(dest, base_off);
duke@435	3871
duke@435	3872	// Compute the length also, if needed:
duke@435	3873	Node* countx = size;
duke@435	3874	countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
duke@435	3875	countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
duke@435	3876
duke@435	3877	// Select an appropriate instruction to initialize the range.
duke@435	3878	// The CopyArray instruction (if supported) can be optimized
duke@435	3879	// into a discrete set of scalar loads and stores.
duke@435	3880	bool disjoint_bases = true;
duke@435	3881	generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
duke@435	3882	src, NULL, dest, NULL, countx);
duke@435	3883
duke@435	3884	// Now that the object is properly initialized, type it as an oop.
duke@435	3885	// Use a secondary InitializeNode memory barrier.
duke@435	3886	InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, raw_adr_idx,
duke@435	3887	raw_obj)->as_Initialize();
duke@435	3888	init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
duke@435	3889	Node* new_obj = new(C, 2) CheckCastPPNode(control(), raw_obj,
duke@435	3890	TypeInstPtr::NOTNULL);
duke@435	3891	new_obj = _gvn.transform(new_obj);
duke@435	3892
duke@435	3893	// If necessary, emit some card marks afterwards. (Non-arrays only.)
duke@435	3894	if (card_mark) {
duke@435	3895	Node* no_particular_value = NULL;
duke@435	3896	Node* no_particular_field = NULL;
duke@435	3897	post_barrier(control(),
duke@435	3898	memory(raw_adr_type),
duke@435	3899	new_obj,
duke@435	3900	no_particular_field,
duke@435	3901	raw_adr_idx,
duke@435	3902	no_particular_value,
duke@435	3903	T_OBJECT,
duke@435	3904	false);
duke@435	3905	}
duke@435	3906	// Present the results of the slow call.
duke@435	3907	result_reg->init_req(_fast_path, control());
duke@435	3908	result_val->init_req(_fast_path, new_obj);
duke@435	3909	result_i_o ->set_req(_fast_path, i_o());
duke@435	3910	result_mem ->set_req(_fast_path, reset_memory());
duke@435	3911	}
duke@435	3912
duke@435	3913	// Return the combined state.
duke@435	3914	set_control( _gvn.transform(result_reg) );
duke@435	3915	set_i_o( _gvn.transform(result_i_o) );
duke@435	3916	set_all_memory( _gvn.transform(result_mem) );
duke@435	3917
duke@435	3918	// Cast the result to a sharper type, since we know what clone does.
duke@435	3919	Node* new_obj = _gvn.transform(result_val);
duke@435	3920	Node* cast = new (C, 2) CheckCastPPNode(control(), new_obj, toop);
duke@435	3921	push(_gvn.transform(cast));
duke@435	3922
duke@435	3923	return true;
duke@435	3924	}
duke@435	3925
duke@435	3926
duke@435	3927	// constants for computing the copy function
duke@435	3928	enum {
duke@435	3929	COPYFUNC_UNALIGNED = 0,
duke@435	3930	COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
duke@435	3931	COPYFUNC_CONJOINT = 0,
duke@435	3932	COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
duke@435	3933	};
duke@435	3934
duke@435	3935	// Note: The condition "disjoint" applies also for overlapping copies
duke@435	3936	// where an descending copy is permitted (i.e., dest_offset <= src_offset).
duke@435	3937	static address
duke@435	3938	select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
duke@435	3939	int selector =
duke@435	3940	(aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
duke@435	3941	(disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
duke@435	3942
duke@435	3943	#define RETURN_STUB(xxx_arraycopy) { \
duke@435	3944	name = #xxx_arraycopy; \
duke@435	3945	return StubRoutines::xxx_arraycopy(); }
duke@435	3946
duke@435	3947	switch (t) {
duke@435	3948	case T_BYTE:
duke@435	3949	case T_BOOLEAN:
duke@435	3950	switch (selector) {
duke@435	3951	case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
duke@435	3952	case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
duke@435	3953	case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
duke@435	3954	case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
duke@435	3955	}
duke@435	3956	case T_CHAR:
duke@435	3957	case T_SHORT:
duke@435	3958	switch (selector) {
duke@435	3959	case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
duke@435	3960	case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
duke@435	3961	case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
duke@435	3962	case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
duke@435	3963	}
duke@435	3964	case T_INT:
duke@435	3965	case T_FLOAT:
duke@435	3966	switch (selector) {
duke@435	3967	case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
duke@435	3968	case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
duke@435	3969	case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
duke@435	3970	case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
duke@435	3971	}
duke@435	3972	case T_DOUBLE:
duke@435	3973	case T_LONG:
duke@435	3974	switch (selector) {
duke@435	3975	case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
duke@435	3976	case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
duke@435	3977	case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
duke@435	3978	case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
duke@435	3979	}
duke@435	3980	case T_ARRAY:
duke@435	3981	case T_OBJECT:
duke@435	3982	switch (selector) {
duke@435	3983	case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
duke@435	3984	case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
duke@435	3985	case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
duke@435	3986	case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
duke@435	3987	}
duke@435	3988	default:
duke@435	3989	ShouldNotReachHere();
duke@435	3990	return NULL;
duke@435	3991	}
duke@435	3992
duke@435	3993	#undef RETURN_STUB
duke@435	3994	}
duke@435	3995
duke@435	3996	//------------------------------basictype2arraycopy----------------------------
duke@435	3997	address LibraryCallKit::basictype2arraycopy(BasicType t,
duke@435	3998	Node* src_offset,
duke@435	3999	Node* dest_offset,
duke@435	4000	bool disjoint_bases,
duke@435	4001	const char* &name) {
duke@435	4002	const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
duke@435	4003	const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
duke@435	4004
duke@435	4005	bool aligned = false;
duke@435	4006	bool disjoint = disjoint_bases;
duke@435	4007
duke@435	4008	// if the offsets are the same, we can treat the memory regions as
duke@435	4009	// disjoint, because either the memory regions are in different arrays,
duke@435	4010	// or they are identical (which we can treat as disjoint.) We can also
duke@435	4011	// treat a copy with a destination index less that the source index
duke@435	4012	// as disjoint since a low->high copy will work correctly in this case.
duke@435	4013	if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
duke@435	4014	dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
duke@435	4015	// both indices are constants
duke@435	4016	int s_offs = src_offset_inttype->get_con();
duke@435	4017	int d_offs = dest_offset_inttype->get_con();
kvn@464	4018	int element_size = type2aelembytes(t);
duke@435	4019	aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
duke@435	4020	((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
duke@435	4021	if (s_offs >= d_offs) disjoint = true;
duke@435	4022	} else if (src_offset == dest_offset && src_offset != NULL) {
duke@435	4023	// This can occur if the offsets are identical non-constants.
duke@435	4024	disjoint = true;
duke@435	4025	}
duke@435	4026
duke@435	4027	return select_arraycopy_function(t, aligned, disjoint, name);
duke@435	4028	}
duke@435	4029
duke@435	4030
duke@435	4031	//------------------------------inline_arraycopy-----------------------
duke@435	4032	bool LibraryCallKit::inline_arraycopy() {
duke@435	4033	// Restore the stack and pop off the arguments.
duke@435	4034	int nargs = 5; // 2 oops, 3 ints, no size_t or long
duke@435	4035	assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
duke@435	4036
duke@435	4037	Node *src = argument(0);
duke@435	4038	Node *src_offset = argument(1);
duke@435	4039	Node *dest = argument(2);
duke@435	4040	Node *dest_offset = argument(3);
duke@435	4041	Node *length = argument(4);
duke@435	4042
duke@435	4043	// Compile time checks. If any of these checks cannot be verified at compile time,
duke@435	4044	// we do not make a fast path for this call. Instead, we let the call remain as it
duke@435	4045	// is. The checks we choose to mandate at compile time are:
duke@435	4046	//
duke@435	4047	// (1) src and dest are arrays.
duke@435	4048	const Type* src_type = src->Value(&_gvn);
duke@435	4049	const Type* dest_type = dest->Value(&_gvn);
duke@435	4050	const TypeAryPtr* top_src = src_type->isa_aryptr();
duke@435	4051	const TypeAryPtr* top_dest = dest_type->isa_aryptr();
duke@435	4052	if (top_src == NULL || top_src->klass() == NULL ||
duke@435	4053	top_dest == NULL || top_dest->klass() == NULL) {
duke@435	4054	// Conservatively insert a memory barrier on all memory slices.
duke@435	4055	// Do not let writes into the source float below the arraycopy.
duke@435	4056	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	4057
duke@435	4058	// Call StubRoutines::generic_arraycopy stub.
duke@435	4059	generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
duke@435	4060	src, src_offset, dest, dest_offset, length,
duke@435	4061	nargs);
duke@435	4062
duke@435	4063	// Do not let reads from the destination float above the arraycopy.
duke@435	4064	// Since we cannot type the arrays, we don't know which slices
duke@435	4065	// might be affected. We could restrict this barrier only to those
duke@435	4066	// memory slices which pertain to array elements--but don't bother.
duke@435	4067	if (!InsertMemBarAfterArraycopy)
duke@435	4068	// (If InsertMemBarAfterArraycopy, there is already one in place.)
duke@435	4069	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	4070	return true;
duke@435	4071	}
duke@435	4072
duke@435	4073	// (2) src and dest arrays must have elements of the same BasicType
duke@435	4074	// Figure out the size and type of the elements we will be copying.
duke@435	4075	BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
duke@435	4076	BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
duke@435	4077	if (src_elem == T_ARRAY) src_elem = T_OBJECT;
duke@435	4078	if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
duke@435	4079
duke@435	4080	if (src_elem != dest_elem || dest_elem == T_VOID) {
duke@435	4081	// The component types are not the same or are not recognized. Punt.
duke@435	4082	// (But, avoid the native method wrapper to JVM_ArrayCopy.)
duke@435	4083	generate_slow_arraycopy(TypePtr::BOTTOM,
duke@435	4084	src, src_offset, dest, dest_offset, length,
duke@435	4085	nargs);
duke@435	4086	return true;
duke@435	4087	}
duke@435	4088
duke@435	4089	//---------------------------------------------------------------------------
duke@435	4090	// We will make a fast path for this call to arraycopy.
duke@435	4091
duke@435	4092	// We have the following tests left to perform:
duke@435	4093	//
duke@435	4094	// (3) src and dest must not be null.
duke@435	4095	// (4) src_offset must not be negative.
duke@435	4096	// (5) dest_offset must not be negative.
duke@435	4097	// (6) length must not be negative.
duke@435	4098	// (7) src_offset + length must not exceed length of src.
duke@435	4099	// (8) dest_offset + length must not exceed length of dest.
duke@435	4100	// (9) each element of an oop array must be assignable
duke@435	4101
duke@435	4102	RegionNode* slow_region = new (C, 1) RegionNode(1);
duke@435	4103	record_for_igvn(slow_region);
duke@435	4104
duke@435	4105	// (3) operands must not be null
duke@435	4106	// We currently perform our null checks with the do_null_check routine.
duke@435	4107	// This means that the null exceptions will be reported in the caller
duke@435	4108	// rather than (correctly) reported inside of the native arraycopy call.
duke@435	4109	// This should be corrected, given time. We do our null check with the
duke@435	4110	// stack pointer restored.
duke@435	4111	_sp += nargs;
duke@435	4112	src = do_null_check(src, T_ARRAY);
duke@435	4113	dest = do_null_check(dest, T_ARRAY);
duke@435	4114	_sp -= nargs;
duke@435	4115
duke@435	4116	// (4) src_offset must not be negative.
duke@435	4117	generate_negative_guard(src_offset, slow_region);
duke@435	4118
duke@435	4119	// (5) dest_offset must not be negative.
duke@435	4120	generate_negative_guard(dest_offset, slow_region);
duke@435	4121
duke@435	4122	// (6) length must not be negative (moved to generate_arraycopy()).
duke@435	4123	// generate_negative_guard(length, slow_region);
duke@435	4124
duke@435	4125	// (7) src_offset + length must not exceed length of src.
duke@435	4126	generate_limit_guard(src_offset, length,
duke@435	4127	load_array_length(src),
duke@435	4128	slow_region);
duke@435	4129
duke@435	4130	// (8) dest_offset + length must not exceed length of dest.
duke@435	4131	generate_limit_guard(dest_offset, length,
duke@435	4132	load_array_length(dest),
duke@435	4133	slow_region);
duke@435	4134
duke@435	4135	// (9) each element of an oop array must be assignable
duke@435	4136	// The generate_arraycopy subroutine checks this.
duke@435	4137
duke@435	4138	// This is where the memory effects are placed:
duke@435	4139	const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
duke@435	4140	generate_arraycopy(adr_type, dest_elem,
duke@435	4141	src, src_offset, dest, dest_offset, length,
duke@435	4142	nargs, false, false, slow_region);
duke@435	4143
duke@435	4144	return true;
duke@435	4145	}
duke@435	4146
duke@435	4147	//-----------------------------generate_arraycopy----------------------
duke@435	4148	// Generate an optimized call to arraycopy.
duke@435	4149	// Caller must guard against non-arrays.
duke@435	4150	// Caller must determine a common array basic-type for both arrays.
duke@435	4151	// Caller must validate offsets against array bounds.
duke@435	4152	// The slow_region has already collected guard failure paths
duke@435	4153	// (such as out of bounds length or non-conformable array types).
duke@435	4154	// The generated code has this shape, in general:
duke@435	4155	//
duke@435	4156	// if (length == 0) return // via zero_path
duke@435	4157	// slowval = -1
duke@435	4158	// if (types unknown) {
duke@435	4159	// slowval = call generic copy loop
duke@435	4160	// if (slowval == 0) return // via checked_path
duke@435	4161	// } else if (indexes in bounds) {
duke@435	4162	// if ((is object array) && !(array type check)) {
duke@435	4163	// slowval = call checked copy loop
duke@435	4164	// if (slowval == 0) return // via checked_path
duke@435	4165	// } else {
duke@435	4166	// call bulk copy loop
duke@435	4167	// return // via fast_path
duke@435	4168	// }
duke@435	4169	// }
duke@435	4170	// // adjust params for remaining work:
duke@435	4171	// if (slowval != -1) {
duke@435	4172	// n = -1^slowval; src_offset += n; dest_offset += n; length -= n
duke@435	4173	// }
duke@435	4174	// slow_region:
duke@435	4175	// call slow arraycopy(src, src_offset, dest, dest_offset, length)
duke@435	4176	// return // via slow_call_path
duke@435	4177	//
duke@435	4178	// This routine is used from several intrinsics: System.arraycopy,
duke@435	4179	// Object.clone (the array subcase), and Arrays.copyOf[Range].
duke@435	4180	//
duke@435	4181	void
duke@435	4182	LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
duke@435	4183	BasicType basic_elem_type,
duke@435	4184	Node* src, Node* src_offset,
duke@435	4185	Node* dest, Node* dest_offset,
duke@435	4186	Node* copy_length,
duke@435	4187	int nargs,
duke@435	4188	bool disjoint_bases,
duke@435	4189	bool length_never_negative,
duke@435	4190	RegionNode* slow_region) {
duke@435	4191
duke@435	4192	if (slow_region == NULL) {
duke@435	4193	slow_region = new(C,1) RegionNode(1);
duke@435	4194	record_for_igvn(slow_region);
duke@435	4195	}
duke@435	4196
duke@435	4197	Node* original_dest = dest;
duke@435	4198	AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
duke@435	4199	Node* raw_dest = NULL; // used before zeroing, if needed
duke@435	4200	bool must_clear_dest = false;
duke@435	4201
duke@435	4202	// See if this is the initialization of a newly-allocated array.
duke@435	4203	// If so, we will take responsibility here for initializing it to zero.
duke@435	4204	// (Note: Because tightly_coupled_allocation performs checks on the
duke@435	4205	// out-edges of the dest, we need to avoid making derived pointers
duke@435	4206	// from it until we have checked its uses.)
duke@435	4207	if (ReduceBulkZeroing
duke@435	4208	&& !ZeroTLAB // pointless if already zeroed
duke@435	4209	&& basic_elem_type != T_CONFLICT // avoid corner case
duke@435	4210	&& !_gvn.eqv_uncast(src, dest)
duke@435	4211	&& ((alloc = tightly_coupled_allocation(dest, slow_region))
duke@435	4212	!= NULL)
kvn@469	4213	&& _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
duke@435	4214	&& alloc->maybe_set_complete(&_gvn)) {
duke@435	4215	// "You break it, you buy it."
duke@435	4216	InitializeNode* init = alloc->initialization();
duke@435	4217	assert(init->is_complete(), "we just did this");
duke@435	4218	assert(dest->Opcode() == Op_CheckCastPP, "sanity");
duke@435	4219	assert(dest->in(0)->in(0) == init, "dest pinned");
duke@435	4220	raw_dest = dest->in(1); // grab the raw pointer!
duke@435	4221	original_dest = dest;
duke@435	4222	dest = raw_dest;
duke@435	4223	adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
duke@435	4224	// Decouple the original InitializeNode, turning it into a simple membar.
duke@435	4225	// We will build a new one at the end of this routine.
duke@435	4226	init->set_req(InitializeNode::RawAddress, top());
duke@435	4227	// From this point on, every exit path is responsible for
duke@435	4228	// initializing any non-copied parts of the object to zero.
duke@435	4229	must_clear_dest = true;
duke@435	4230	} else {
duke@435	4231	// No zeroing elimination here.
duke@435	4232	alloc = NULL;
duke@435	4233	//original_dest = dest;
duke@435	4234	//must_clear_dest = false;
duke@435	4235	}
duke@435	4236
duke@435	4237	// Results are placed here:
duke@435	4238	enum { fast_path = 1, // normal void-returning assembly stub
duke@435	4239	checked_path = 2, // special assembly stub with cleanup
duke@435	4240	slow_call_path = 3, // something went wrong; call the VM
duke@435	4241	zero_path = 4, // bypass when length of copy is zero
duke@435	4242	bcopy_path = 5, // copy primitive array by 64-bit blocks
duke@435	4243	PATH_LIMIT = 6
duke@435	4244	};
duke@435	4245	RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	4246	PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
duke@435	4247	PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
duke@435	4248	record_for_igvn(result_region);
duke@435	4249	_gvn.set_type_bottom(result_i_o);
duke@435	4250	_gvn.set_type_bottom(result_memory);
duke@435	4251	assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
duke@435	4252
duke@435	4253	// The slow_control path:
duke@435	4254	Node* slow_control;
duke@435	4255	Node* slow_i_o = i_o();
duke@435	4256	Node* slow_mem = memory(adr_type);
duke@435	4257	debug_only(slow_control = (Node*) badAddress);
duke@435	4258
duke@435	4259	// Checked control path:
duke@435	4260	Node* checked_control = top();
duke@435	4261	Node* checked_mem = NULL;
duke@435	4262	Node* checked_i_o = NULL;
duke@435	4263	Node* checked_value = NULL;
duke@435	4264
duke@435	4265	if (basic_elem_type == T_CONFLICT) {
duke@435	4266	assert(!must_clear_dest, "");
duke@435	4267	Node* cv = generate_generic_arraycopy(adr_type,
duke@435	4268	src, src_offset, dest, dest_offset,
duke@435	4269	copy_length, nargs);
duke@435	4270	if (cv == NULL) cv = intcon(-1); // failure (no stub available)
duke@435	4271	checked_control = control();
duke@435	4272	checked_i_o = i_o();
duke@435	4273	checked_mem = memory(adr_type);
duke@435	4274	checked_value = cv;
duke@435	4275	set_control(top()); // no fast path
duke@435	4276	}
duke@435	4277
duke@435	4278	Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
duke@435	4279	if (not_pos != NULL) {
duke@435	4280	PreserveJVMState pjvms(this);
duke@435	4281	set_control(not_pos);
duke@435	4282
duke@435	4283	// (6) length must not be negative.
duke@435	4284	if (!length_never_negative) {
duke@435	4285	generate_negative_guard(copy_length, slow_region);
duke@435	4286	}
duke@435	4287
duke@435	4288	if (!stopped() && must_clear_dest) {
duke@435	4289	Node* dest_length = alloc->in(AllocateNode::ALength);
duke@435	4290	if (_gvn.eqv_uncast(copy_length, dest_length)
duke@435	4291	|| _gvn.find_int_con(dest_length, 1) <= 0) {
duke@435	4292	// There is no zeroing to do.
duke@435	4293	} else {
duke@435	4294	// Clear the whole thing since there are no source elements to copy.
duke@435	4295	generate_clear_array(adr_type, dest, basic_elem_type,
duke@435	4296	intcon(0), NULL,
duke@435	4297	alloc->in(AllocateNode::AllocSize));
duke@435	4298	}
duke@435	4299	}
duke@435	4300
duke@435	4301	// Present the results of the fast call.
duke@435	4302	result_region->init_req(zero_path, control());
duke@435	4303	result_i_o ->init_req(zero_path, i_o());
duke@435	4304	result_memory->init_req(zero_path, memory(adr_type));
duke@435	4305	}
duke@435	4306
duke@435	4307	if (!stopped() && must_clear_dest) {
duke@435	4308	// We have to initialize the *uncopied* part of the array to zero.
duke@435	4309	// The copy destination is the slice dest[off..off+len]. The other slices
duke@435	4310	// are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
duke@435	4311	Node* dest_size = alloc->in(AllocateNode::AllocSize);
duke@435	4312	Node* dest_length = alloc->in(AllocateNode::ALength);
duke@435	4313	Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
duke@435	4314	copy_length) );
duke@435	4315
duke@435	4316	// If there is a head section that needs zeroing, do it now.
duke@435	4317	if (find_int_con(dest_offset, -1) != 0) {
duke@435	4318	generate_clear_array(adr_type, dest, basic_elem_type,
duke@435	4319	intcon(0), dest_offset,
duke@435	4320	NULL);
duke@435	4321	}
duke@435	4322
duke@435	4323	// Next, perform a dynamic check on the tail length.
duke@435	4324	// It is often zero, and we can win big if we prove this.
duke@435	4325	// There are two wins: Avoid generating the ClearArray
duke@435	4326	// with its attendant messy index arithmetic, and upgrade
duke@435	4327	// the copy to a more hardware-friendly word size of 64 bits.
duke@435	4328	Node* tail_ctl = NULL;
duke@435	4329	if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
duke@435	4330	Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
duke@435	4331	Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
duke@435	4332	tail_ctl = generate_slow_guard(bol_lt, NULL);
duke@435	4333	assert(tail_ctl != NULL || !stopped(), "must be an outcome");
duke@435	4334	}
duke@435	4335
duke@435	4336	// At this point, let's assume there is no tail.
duke@435	4337	if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
duke@435	4338	// There is no tail. Try an upgrade to a 64-bit copy.
duke@435	4339	bool didit = false;
duke@435	4340	{ PreserveJVMState pjvms(this);
duke@435	4341	didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
duke@435	4342	src, src_offset, dest, dest_offset,
duke@435	4343	dest_size);
duke@435	4344	if (didit) {
duke@435	4345	// Present the results of the block-copying fast call.
duke@435	4346	result_region->init_req(bcopy_path, control());
duke@435	4347	result_i_o ->init_req(bcopy_path, i_o());
duke@435	4348	result_memory->init_req(bcopy_path, memory(adr_type));
duke@435	4349	}
duke@435	4350	}
duke@435	4351	if (didit)
duke@435	4352	set_control(top()); // no regular fast path
duke@435	4353	}
duke@435	4354
duke@435	4355	// Clear the tail, if any.
duke@435	4356	if (tail_ctl != NULL) {
duke@435	4357	Node* notail_ctl = stopped() ? NULL : control();
duke@435	4358	set_control(tail_ctl);
duke@435	4359	if (notail_ctl == NULL) {
duke@435	4360	generate_clear_array(adr_type, dest, basic_elem_type,
duke@435	4361	dest_tail, NULL,
duke@435	4362	dest_size);
duke@435	4363	} else {
duke@435	4364	// Make a local merge.
duke@435	4365	Node* done_ctl = new(C,3) RegionNode(3);
duke@435	4366	Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
duke@435	4367	done_ctl->init_req(1, notail_ctl);
duke@435	4368	done_mem->init_req(1, memory(adr_type));
duke@435	4369	generate_clear_array(adr_type, dest, basic_elem_type,
duke@435	4370	dest_tail, NULL,
duke@435	4371	dest_size);
duke@435	4372	done_ctl->init_req(2, control());
duke@435	4373	done_mem->init_req(2, memory(adr_type));
duke@435	4374	set_control( _gvn.transform(done_ctl) );
duke@435	4375	set_memory( _gvn.transform(done_mem), adr_type );
duke@435	4376	}
duke@435	4377	}
duke@435	4378	}
duke@435	4379
duke@435	4380	BasicType copy_type = basic_elem_type;
duke@435	4381	assert(basic_elem_type != T_ARRAY, "caller must fix this");
duke@435	4382	if (!stopped() && copy_type == T_OBJECT) {
duke@435	4383	// If src and dest have compatible element types, we can copy bits.
duke@435	4384	// Types S[] and D[] are compatible if D is a supertype of S.
duke@435	4385	//
duke@435	4386	// If they are not, we will use checked_oop_disjoint_arraycopy,
duke@435	4387	// which performs a fast optimistic per-oop check, and backs off
duke@435	4388	// further to JVM_ArrayCopy on the first per-oop check that fails.
duke@435	4389	// (Actually, we don't move raw bits only; the GC requires card marks.)
duke@435	4390
duke@435	4391	// Get the klassOop for both src and dest
duke@435	4392	Node* src_klass = load_object_klass(src);
duke@435	4393	Node* dest_klass = load_object_klass(dest);
duke@435	4394
duke@435	4395	// Generate the subtype check.
duke@435	4396	// This might fold up statically, or then again it might not.
duke@435	4397	//
duke@435	4398	// Non-static example: Copying List<String>.elements to a new String[].
duke@435	4399	// The backing store for a List<String> is always an Object[],
duke@435	4400	// but its elements are always type String, if the generic types
duke@435	4401	// are correct at the source level.
duke@435	4402	//
duke@435	4403	// Test S[] against D[], not S against D, because (probably)
duke@435	4404	// the secondary supertype cache is less busy for S[] than S.
duke@435	4405	// This usually only matters when D is an interface.
duke@435	4406	Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
duke@435	4407	// Plug failing path into checked_oop_disjoint_arraycopy
duke@435	4408	if (not_subtype_ctrl != top()) {
duke@435	4409	PreserveJVMState pjvms(this);
duke@435	4410	set_control(not_subtype_ctrl);
duke@435	4411	// (At this point we can assume disjoint_bases, since types differ.)
duke@435	4412	int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
duke@435	4413	Node* p1 = basic_plus_adr(dest_klass, ek_offset);
kvn@599	4414	Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
duke@435	4415	Node* dest_elem_klass = _gvn.transform(n1);
duke@435	4416	Node* cv = generate_checkcast_arraycopy(adr_type,
duke@435	4417	dest_elem_klass,
duke@435	4418	src, src_offset, dest, dest_offset,
duke@435	4419	copy_length,
duke@435	4420	nargs);
duke@435	4421	if (cv == NULL) cv = intcon(-1); // failure (no stub available)
duke@435	4422	checked_control = control();
duke@435	4423	checked_i_o = i_o();
duke@435	4424	checked_mem = memory(adr_type);
duke@435	4425	checked_value = cv;
duke@435	4426	}
duke@435	4427	// At this point we know we do not need type checks on oop stores.
duke@435	4428
duke@435	4429	// Let's see if we need card marks:
duke@435	4430	if (alloc != NULL && use_ReduceInitialCardMarks()) {
duke@435	4431	// If we do not need card marks, copy using the jint or jlong stub.
coleenp@548	4432	copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
kvn@464	4433	assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
duke@435	4434	"sizes agree");
duke@435	4435	}
duke@435	4436	}
duke@435	4437
duke@435	4438	if (!stopped()) {
duke@435	4439	// Generate the fast path, if possible.
duke@435	4440	PreserveJVMState pjvms(this);
duke@435	4441	generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
duke@435	4442	src, src_offset, dest, dest_offset,
duke@435	4443	ConvI2X(copy_length));
duke@435	4444
duke@435	4445	// Present the results of the fast call.
duke@435	4446	result_region->init_req(fast_path, control());
duke@435	4447	result_i_o ->init_req(fast_path, i_o());
duke@435	4448	result_memory->init_req(fast_path, memory(adr_type));
duke@435	4449	}
duke@435	4450
duke@435	4451	// Here are all the slow paths up to this point, in one bundle:
duke@435	4452	slow_control = top();
duke@435	4453	if (slow_region != NULL)
duke@435	4454	slow_control = _gvn.transform(slow_region);
duke@435	4455	debug_only(slow_region = (RegionNode*)badAddress);
duke@435	4456
duke@435	4457	set_control(checked_control);
duke@435	4458	if (!stopped()) {
duke@435	4459	// Clean up after the checked call.
duke@435	4460	// The returned value is either 0 or -1^K,
duke@435	4461	// where K = number of partially transferred array elements.
duke@435	4462	Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
duke@435	4463	Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
duke@435	4464	IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
duke@435	4465
duke@435	4466	// If it is 0, we are done, so transfer to the end.
duke@435	4467	Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
duke@435	4468	result_region->init_req(checked_path, checks_done);
duke@435	4469	result_i_o ->init_req(checked_path, checked_i_o);
duke@435	4470	result_memory->init_req(checked_path, checked_mem);
duke@435	4471
duke@435	4472	// If it is not zero, merge into the slow call.
duke@435	4473	set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
duke@435	4474	RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
duke@435	4475	PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
duke@435	4476	PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
duke@435	4477	record_for_igvn(slow_reg2);
duke@435	4478	slow_reg2 ->init_req(1, slow_control);
duke@435	4479	slow_i_o2 ->init_req(1, slow_i_o);
duke@435	4480	slow_mem2 ->init_req(1, slow_mem);
duke@435	4481	slow_reg2 ->init_req(2, control());
duke@435	4482	slow_i_o2 ->init_req(2, i_o());
duke@435	4483	slow_mem2 ->init_req(2, memory(adr_type));
duke@435	4484
duke@435	4485	slow_control = _gvn.transform(slow_reg2);
duke@435	4486	slow_i_o = _gvn.transform(slow_i_o2);
duke@435	4487	slow_mem = _gvn.transform(slow_mem2);
duke@435	4488
duke@435	4489	if (alloc != NULL) {
duke@435	4490	// We'll restart from the very beginning, after zeroing the whole thing.
duke@435	4491	// This can cause double writes, but that's OK since dest is brand new.
duke@435	4492	// So we ignore the low 31 bits of the value returned from the stub.
duke@435	4493	} else {
duke@435	4494	// We must continue the copy exactly where it failed, or else
duke@435	4495	// another thread might see the wrong number of writes to dest.
duke@435	4496	Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
duke@435	4497	Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
duke@435	4498	slow_offset->init_req(1, intcon(0));
duke@435	4499	slow_offset->init_req(2, checked_offset);
duke@435	4500	slow_offset = _gvn.transform(slow_offset);
duke@435	4501
duke@435	4502	// Adjust the arguments by the conditionally incoming offset.
duke@435	4503	Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
duke@435	4504	Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
duke@435	4505	Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
duke@435	4506
duke@435	4507	// Tweak the node variables to adjust the code produced below:
duke@435	4508	src_offset = src_off_plus;
duke@435	4509	dest_offset = dest_off_plus;
duke@435	4510	copy_length = length_minus;
duke@435	4511	}
duke@435	4512	}
duke@435	4513
duke@435	4514	set_control(slow_control);
duke@435	4515	if (!stopped()) {
duke@435	4516	// Generate the slow path, if needed.
duke@435	4517	PreserveJVMState pjvms(this); // replace_in_map may trash the map
duke@435	4518
duke@435	4519	set_memory(slow_mem, adr_type);
duke@435	4520	set_i_o(slow_i_o);
duke@435	4521
duke@435	4522	if (must_clear_dest) {
duke@435	4523	generate_clear_array(adr_type, dest, basic_elem_type,
duke@435	4524	intcon(0), NULL,
duke@435	4525	alloc->in(AllocateNode::AllocSize));
duke@435	4526	}
duke@435	4527
duke@435	4528	if (dest != original_dest) {
duke@435	4529	// Promote from rawptr to oop, so it looks right in the call's GC map.
duke@435	4530	dest = _gvn.transform( new(C,2) CheckCastPPNode(control(), dest,
duke@435	4531	TypeInstPtr::NOTNULL) );
duke@435	4532
duke@435	4533	// Edit the call's debug-info to avoid referring to original_dest.
duke@435	4534	// (The problem with original_dest is that it isn't ready until
duke@435	4535	// after the InitializeNode completes, but this stuff is before.)
duke@435	4536	// Substitute in the locally valid dest_oop.
duke@435	4537	replace_in_map(original_dest, dest);
duke@435	4538	}
duke@435	4539
duke@435	4540	generate_slow_arraycopy(adr_type,
duke@435	4541	src, src_offset, dest, dest_offset,
duke@435	4542	copy_length, nargs);
duke@435	4543
duke@435	4544	result_region->init_req(slow_call_path, control());
duke@435	4545	result_i_o ->init_req(slow_call_path, i_o());
duke@435	4546	result_memory->init_req(slow_call_path, memory(adr_type));
duke@435	4547	}
duke@435	4548
duke@435	4549	// Remove unused edges.
duke@435	4550	for (uint i = 1; i < result_region->req(); i++) {
duke@435	4551	if (result_region->in(i) == NULL)
duke@435	4552	result_region->init_req(i, top());
duke@435	4553	}
duke@435	4554
duke@435	4555	// Finished; return the combined state.
duke@435	4556	set_control( _gvn.transform(result_region) );
duke@435	4557	set_i_o( _gvn.transform(result_i_o) );
duke@435	4558	set_memory( _gvn.transform(result_memory), adr_type );
duke@435	4559
duke@435	4560	if (dest != original_dest) {
duke@435	4561	// Pin the "finished" array node after the arraycopy/zeroing operations.
duke@435	4562	// Use a secondary InitializeNode memory barrier.
duke@435	4563	InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
duke@435	4564	Compile::AliasIdxRaw,
duke@435	4565	raw_dest)->as_Initialize();
duke@435	4566	init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
duke@435	4567	_gvn.hash_delete(original_dest);
duke@435	4568	original_dest->set_req(0, control());
duke@435	4569	_gvn.hash_find_insert(original_dest); // put back into GVN table
duke@435	4570	}
duke@435	4571
duke@435	4572	// The memory edges above are precise in order to model effects around
duke@435	4573	// array copyies accurately to allow value numbering of field loads around
duke@435	4574	// arraycopy. Such field loads, both before and after, are common in Java
duke@435	4575	// collections and similar classes involving header/array data structures.
duke@435	4576	//
duke@435	4577	// But with low number of register or when some registers are used or killed
duke@435	4578	// by arraycopy calls it causes registers spilling on stack. See 6544710.
duke@435	4579	// The next memory barrier is added to avoid it. If the arraycopy can be
duke@435	4580	// optimized away (which it can, sometimes) then we can manually remove
duke@435	4581	// the membar also.
duke@435	4582	if (InsertMemBarAfterArraycopy)
duke@435	4583	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	4584	}
duke@435	4585
duke@435	4586
duke@435	4587	// Helper function which determines if an arraycopy immediately follows
duke@435	4588	// an allocation, with no intervening tests or other escapes for the object.
duke@435	4589	AllocateArrayNode*
duke@435	4590	LibraryCallKit::tightly_coupled_allocation(Node* ptr,
duke@435	4591	RegionNode* slow_region) {
duke@435	4592	if (stopped()) return NULL; // no fast path
duke@435	4593	if (C->AliasLevel() == 0) return NULL; // no MergeMems around
duke@435	4594
duke@435	4595	AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
duke@435	4596	if (alloc == NULL) return NULL;
duke@435	4597
duke@435	4598	Node* rawmem = memory(Compile::AliasIdxRaw);
duke@435	4599	// Is the allocation's memory state untouched?
duke@435	4600	if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
duke@435	4601	// Bail out if there have been raw-memory effects since the allocation.
duke@435	4602	// (Example: There might have been a call or safepoint.)
duke@435	4603	return NULL;
duke@435	4604	}
duke@435	4605	rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
duke@435	4606	if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
duke@435	4607	return NULL;
duke@435	4608	}
duke@435	4609
duke@435	4610	// There must be no unexpected observers of this allocation.
duke@435	4611	for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
duke@435	4612	Node* obs = ptr->fast_out(i);
duke@435	4613	if (obs != this->map()) {
duke@435	4614	return NULL;
duke@435	4615	}
duke@435	4616	}
duke@435	4617
duke@435	4618	// This arraycopy must unconditionally follow the allocation of the ptr.
duke@435	4619	Node* alloc_ctl = ptr->in(0);
duke@435	4620	assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
duke@435	4621
duke@435	4622	Node* ctl = control();
duke@435	4623	while (ctl != alloc_ctl) {
duke@435	4624	// There may be guards which feed into the slow_region.
duke@435	4625	// Any other control flow means that we might not get a chance
duke@435	4626	// to finish initializing the allocated object.
duke@435	4627	if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
duke@435	4628	IfNode* iff = ctl->in(0)->as_If();
duke@435	4629	Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
duke@435	4630	assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
duke@435	4631	if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
duke@435	4632	ctl = iff->in(0); // This test feeds the known slow_region.
duke@435	4633	continue;
duke@435	4634	}
duke@435	4635	// One more try: Various low-level checks bottom out in
duke@435	4636	// uncommon traps. If the debug-info of the trap omits
duke@435	4637	// any reference to the allocation, as we've already
duke@435	4638	// observed, then there can be no objection to the trap.
duke@435	4639	bool found_trap = false;
duke@435	4640	for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
duke@435	4641	Node* obs = not_ctl->fast_out(j);
duke@435	4642	if (obs->in(0) == not_ctl && obs->is_Call() &&
duke@435	4643	(obs->as_Call()->entry_point() ==
duke@435	4644	SharedRuntime::uncommon_trap_blob()->instructions_begin())) {
duke@435	4645	found_trap = true; break;
duke@435	4646	}
duke@435	4647	}
duke@435	4648	if (found_trap) {
duke@435	4649	ctl = iff->in(0); // This test feeds a harmless uncommon trap.
duke@435	4650	continue;
duke@435	4651	}
duke@435	4652	}
duke@435	4653	return NULL;
duke@435	4654	}
duke@435	4655
duke@435	4656	// If we get this far, we have an allocation which immediately
duke@435	4657	// precedes the arraycopy, and we can take over zeroing the new object.
duke@435	4658	// The arraycopy will finish the initialization, and provide
duke@435	4659	// a new control state to which we will anchor the destination pointer.
duke@435	4660
duke@435	4661	return alloc;
duke@435	4662	}
duke@435	4663
duke@435	4664	// Helper for initialization of arrays, creating a ClearArray.
duke@435	4665	// It writes zero bits in [start..end), within the body of an array object.
duke@435	4666	// The memory effects are all chained onto the 'adr_type' alias category.
duke@435	4667	//
duke@435	4668	// Since the object is otherwise uninitialized, we are free
duke@435	4669	// to put a little "slop" around the edges of the cleared area,
duke@435	4670	// as long as it does not go back into the array's header,
duke@435	4671	// or beyond the array end within the heap.
duke@435	4672	//
duke@435	4673	// The lower edge can be rounded down to the nearest jint and the
duke@435	4674	// upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
duke@435	4675	//
duke@435	4676	// Arguments:
duke@435	4677	// adr_type memory slice where writes are generated
duke@435	4678	// dest oop of the destination array
duke@435	4679	// basic_elem_type element type of the destination
duke@435	4680	// slice_idx array index of first element to store
duke@435	4681	// slice_len number of elements to store (or NULL)
duke@435	4682	// dest_size total size in bytes of the array object
duke@435	4683	//
duke@435	4684	// Exactly one of slice_len or dest_size must be non-NULL.
duke@435	4685	// If dest_size is non-NULL, zeroing extends to the end of the object.
duke@435	4686	// If slice_len is non-NULL, the slice_idx value must be a constant.
duke@435	4687	void
duke@435	4688	LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
duke@435	4689	Node* dest,
duke@435	4690	BasicType basic_elem_type,
duke@435	4691	Node* slice_idx,
duke@435	4692	Node* slice_len,
duke@435	4693	Node* dest_size) {
duke@435	4694	// one or the other but not both of slice_len and dest_size:
duke@435	4695	assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
duke@435	4696	if (slice_len == NULL) slice_len = top();
duke@435	4697	if (dest_size == NULL) dest_size = top();
duke@435	4698
duke@435	4699	// operate on this memory slice:
duke@435	4700	Node* mem = memory(adr_type); // memory slice to operate on
duke@435	4701
duke@435	4702	// scaling and rounding of indexes:
kvn@464	4703	int scale = exact_log2(type2aelembytes(basic_elem_type));
duke@435	4704	int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
duke@435	4705	int clear_low = (-1 << scale) & (BytesPerInt - 1);
duke@435	4706	int bump_bit = (-1 << scale) & BytesPerInt;
duke@435	4707
duke@435	4708	// determine constant starts and ends
duke@435	4709	const intptr_t BIG_NEG = -128;
duke@435	4710	assert(BIG_NEG + 2*abase < 0, "neg enough");
duke@435	4711	intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
duke@435	4712	intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
duke@435	4713	if (slice_len_con == 0) {
duke@435	4714	return; // nothing to do here
duke@435	4715	}
duke@435	4716	intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
duke@435	4717	intptr_t end_con = find_intptr_t_con(dest_size, -1);
duke@435	4718	if (slice_idx_con >= 0 && slice_len_con >= 0) {
duke@435	4719	assert(end_con < 0, "not two cons");
duke@435	4720	end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
duke@435	4721	BytesPerLong);
duke@435	4722	}
duke@435	4723
duke@435	4724	if (start_con >= 0 && end_con >= 0) {
duke@435	4725	// Constant start and end. Simple.
duke@435	4726	mem = ClearArrayNode::clear_memory(control(), mem, dest,
duke@435	4727	start_con, end_con, &_gvn);
duke@435	4728	} else if (start_con >= 0 && dest_size != top()) {
duke@435	4729	// Constant start, pre-rounded end after the tail of the array.
duke@435	4730	Node* end = dest_size;
duke@435	4731	mem = ClearArrayNode::clear_memory(control(), mem, dest,
duke@435	4732	start_con, end, &_gvn);
duke@435	4733	} else if (start_con >= 0 && slice_len != top()) {
duke@435	4734	// Constant start, non-constant end. End needs rounding up.
duke@435	4735	// End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
duke@435	4736	intptr_t end_base = abase + (slice_idx_con << scale);
duke@435	4737	int end_round = (-1 << scale) & (BytesPerLong - 1);
duke@435	4738	Node* end = ConvI2X(slice_len);
duke@435	4739	if (scale != 0)
duke@435	4740	end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
duke@435	4741	end_base += end_round;
duke@435	4742	end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
duke@435	4743	end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
duke@435	4744	mem = ClearArrayNode::clear_memory(control(), mem, dest,
duke@435	4745	start_con, end, &_gvn);
duke@435	4746	} else if (start_con < 0 && dest_size != top()) {
duke@435	4747	// Non-constant start, pre-rounded end after the tail of the array.
duke@435	4748	// This is almost certainly a "round-to-end" operation.
duke@435	4749	Node* start = slice_idx;
duke@435	4750	start = ConvI2X(start);
duke@435	4751	if (scale != 0)
duke@435	4752	start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
duke@435	4753	start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
duke@435	4754	if ((bump_bit | clear_low) != 0) {
duke@435	4755	int to_clear = (bump_bit | clear_low);
duke@435	4756	// Align up mod 8, then store a jint zero unconditionally
duke@435	4757	// just before the mod-8 boundary.
coleenp@548	4758	if (((abase + bump_bit) & ~to_clear) - bump_bit
coleenp@548	4759	< arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
coleenp@548	4760	bump_bit = 0;
coleenp@548	4761	assert((abase & to_clear) == 0, "array base must be long-aligned");
coleenp@548	4762	} else {
coleenp@548	4763	// Bump 'start' up to (or past) the next jint boundary:
coleenp@548	4764	start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
coleenp@548	4765	assert((abase & clear_low) == 0, "array base must be int-aligned");
coleenp@548	4766	}
duke@435	4767	// Round bumped 'start' down to jlong boundary in body of array.
duke@435	4768	start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
coleenp@548	4769	if (bump_bit != 0) {
coleenp@548	4770	// Store a zero to the immediately preceding jint:
coleenp@548	4771	Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
coleenp@548	4772	Node* p1 = basic_plus_adr(dest, x1);
coleenp@548	4773	mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
coleenp@548	4774	mem = _gvn.transform(mem);
coleenp@548	4775	}
duke@435	4776	}
duke@435	4777	Node* end = dest_size; // pre-rounded
duke@435	4778	mem = ClearArrayNode::clear_memory(control(), mem, dest,
duke@435	4779	start, end, &_gvn);
duke@435	4780	} else {
duke@435	4781	// Non-constant start, unrounded non-constant end.
duke@435	4782	// (Nobody zeroes a random midsection of an array using this routine.)
duke@435	4783	ShouldNotReachHere(); // fix caller
duke@435	4784	}
duke@435	4785
duke@435	4786	// Done.
duke@435	4787	set_memory(mem, adr_type);
duke@435	4788	}
duke@435	4789
duke@435	4790
duke@435	4791	bool
duke@435	4792	LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
duke@435	4793	BasicType basic_elem_type,
duke@435	4794	AllocateNode* alloc,
duke@435	4795	Node* src, Node* src_offset,
duke@435	4796	Node* dest, Node* dest_offset,
duke@435	4797	Node* dest_size) {
duke@435	4798	// See if there is an advantage from block transfer.
kvn@464	4799	int scale = exact_log2(type2aelembytes(basic_elem_type));
duke@435	4800	if (scale >= LogBytesPerLong)
duke@435	4801	return false; // it is already a block transfer
duke@435	4802
duke@435	4803	// Look at the alignment of the starting offsets.
duke@435	4804	int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
duke@435	4805	const intptr_t BIG_NEG = -128;
duke@435	4806	assert(BIG_NEG + 2*abase < 0, "neg enough");
duke@435	4807
duke@435	4808	intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
duke@435	4809	intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
duke@435	4810	if (src_off < 0 || dest_off < 0)
duke@435	4811	// At present, we can only understand constants.
duke@435	4812	return false;
duke@435	4813
duke@435	4814	if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
duke@435	4815	// Non-aligned; too bad.
duke@435	4816	// One more chance: Pick off an initial 32-bit word.
duke@435	4817	// This is a common case, since abase can be odd mod 8.
duke@435	4818	if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
duke@435	4819	((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
duke@435	4820	Node* sptr = basic_plus_adr(src, src_off);
duke@435	4821	Node* dptr = basic_plus_adr(dest, dest_off);
duke@435	4822	Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
duke@435	4823	store_to_memory(control(), dptr, sval, T_INT, adr_type);
duke@435	4824	src_off += BytesPerInt;
duke@435	4825	dest_off += BytesPerInt;
duke@435	4826	} else {
duke@435	4827	return false;
duke@435	4828	}
duke@435	4829	}
duke@435	4830	assert(src_off % BytesPerLong == 0, "");
duke@435	4831	assert(dest_off % BytesPerLong == 0, "");
duke@435	4832
duke@435	4833	// Do this copy by giant steps.
duke@435	4834	Node* sptr = basic_plus_adr(src, src_off);
duke@435	4835	Node* dptr = basic_plus_adr(dest, dest_off);
duke@435	4836	Node* countx = dest_size;
duke@435	4837	countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
duke@435	4838	countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
duke@435	4839
duke@435	4840	bool disjoint_bases = true; // since alloc != NULL
duke@435	4841	generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
duke@435	4842	sptr, NULL, dptr, NULL, countx);
duke@435	4843
duke@435	4844	return true;
duke@435	4845	}
duke@435	4846
duke@435	4847
duke@435	4848	// Helper function; generates code for the slow case.
duke@435	4849	// We make a call to a runtime method which emulates the native method,
duke@435	4850	// but without the native wrapper overhead.
duke@435	4851	void
duke@435	4852	LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
duke@435	4853	Node* src, Node* src_offset,
duke@435	4854	Node* dest, Node* dest_offset,
duke@435	4855	Node* copy_length,
duke@435	4856	int nargs) {
duke@435	4857	_sp += nargs; // any deopt will start just before call to enclosing method
duke@435	4858	Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
duke@435	4859	OptoRuntime::slow_arraycopy_Type(),
duke@435	4860	OptoRuntime::slow_arraycopy_Java(),
duke@435	4861	"slow_arraycopy", adr_type,
duke@435	4862	src, src_offset, dest, dest_offset,
duke@435	4863	copy_length);
duke@435	4864	_sp -= nargs;
duke@435	4865
duke@435	4866	// Handle exceptions thrown by this fellow:
duke@435	4867	make_slow_call_ex(call, env()->Throwable_klass(), false);
duke@435	4868	}
duke@435	4869
duke@435	4870	// Helper function; generates code for cases requiring runtime checks.
duke@435	4871	Node*
duke@435	4872	LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
duke@435	4873	Node* dest_elem_klass,
duke@435	4874	Node* src, Node* src_offset,
duke@435	4875	Node* dest, Node* dest_offset,
duke@435	4876	Node* copy_length,
duke@435	4877	int nargs) {
duke@435	4878	if (stopped()) return NULL;
duke@435	4879
duke@435	4880	address copyfunc_addr = StubRoutines::checkcast_arraycopy();
duke@435	4881	if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
duke@435	4882	return NULL;
duke@435	4883	}
duke@435	4884
duke@435	4885	// Pick out the parameters required to perform a store-check
duke@435	4886	// for the target array. This is an optimistic check. It will
duke@435	4887	// look in each non-null element's class, at the desired klass's
duke@435	4888	// super_check_offset, for the desired klass.
duke@435	4889	int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
duke@435	4890	Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
duke@435	4891	Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM);
duke@435	4892	Node* check_offset = _gvn.transform(n3);
duke@435	4893	Node* check_value = dest_elem_klass;
duke@435	4894
duke@435	4895	Node* src_start = array_element_address(src, src_offset, T_OBJECT);
duke@435	4896	Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
duke@435	4897
duke@435	4898	// (We know the arrays are never conjoint, because their types differ.)
duke@435	4899	Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
duke@435	4900	OptoRuntime::checkcast_arraycopy_Type(),
duke@435	4901	copyfunc_addr, "checkcast_arraycopy", adr_type,
duke@435	4902	// five arguments, of which two are
duke@435	4903	// intptr_t (jlong in LP64)
duke@435	4904	src_start, dest_start,
duke@435	4905	copy_length XTOP,
duke@435	4906	check_offset XTOP,
duke@435	4907	check_value);
duke@435	4908
duke@435	4909	return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
duke@435	4910	}
duke@435	4911
duke@435	4912
duke@435	4913	// Helper function; generates code for cases requiring runtime checks.
duke@435	4914	Node*
duke@435	4915	LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
duke@435	4916	Node* src, Node* src_offset,
duke@435	4917	Node* dest, Node* dest_offset,
duke@435	4918	Node* copy_length,
duke@435	4919	int nargs) {
duke@435	4920	if (stopped()) return NULL;
duke@435	4921
duke@435	4922	address copyfunc_addr = StubRoutines::generic_arraycopy();
duke@435	4923	if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
duke@435	4924	return NULL;
duke@435	4925	}
duke@435	4926
duke@435	4927	Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
duke@435	4928	OptoRuntime::generic_arraycopy_Type(),
duke@435	4929	copyfunc_addr, "generic_arraycopy", adr_type,
duke@435	4930	src, src_offset, dest, dest_offset, copy_length);
duke@435	4931
duke@435	4932	return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
duke@435	4933	}
duke@435	4934
duke@435	4935	// Helper function; generates the fast out-of-line call to an arraycopy stub.
duke@435	4936	void
duke@435	4937	LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
duke@435	4938	BasicType basic_elem_type,
duke@435	4939	bool disjoint_bases,
duke@435	4940	Node* src, Node* src_offset,
duke@435	4941	Node* dest, Node* dest_offset,
duke@435	4942	Node* copy_length) {
duke@435	4943	if (stopped()) return; // nothing to do
duke@435	4944
duke@435	4945	Node* src_start = src;
duke@435	4946	Node* dest_start = dest;
duke@435	4947	if (src_offset != NULL || dest_offset != NULL) {
duke@435	4948	assert(src_offset != NULL && dest_offset != NULL, "");
duke@435	4949	src_start = array_element_address(src, src_offset, basic_elem_type);
duke@435	4950	dest_start = array_element_address(dest, dest_offset, basic_elem_type);
duke@435	4951	}
duke@435	4952
duke@435	4953	// Figure out which arraycopy runtime method to call.
duke@435	4954	const char* copyfunc_name = "arraycopy";
duke@435	4955	address copyfunc_addr =
duke@435	4956	basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
duke@435	4957	disjoint_bases, copyfunc_name);
duke@435	4958
duke@435	4959	// Call it. Note that the count_ix value is not scaled to a byte-size.
duke@435	4960	make_runtime_call(RC_LEAF|RC_NO_FP,
duke@435	4961	OptoRuntime::fast_arraycopy_Type(),
duke@435	4962	copyfunc_addr, copyfunc_name, adr_type,
duke@435	4963	src_start, dest_start, copy_length XTOP);
duke@435	4964	}