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

src/share/vm/opto/library_call.cpp@ea3f9723b5cf (annotated)

src/share/vm/opto/library_call.cpp

Thu, 23 Jul 2009 14:53:56 -0700

author

kvn

date

Thu, 23 Jul 2009 14:53:56 -0700

changeset 1294

ea3f9723b5cf

parent 1291

75596850f863

child 1331

1cef5ec3ca56

permissions

-rw-r--r--

6860599: nodes limit could be reached during Output phase
Summary: Bailout compilation if nodes limit could be reached during Output phase.
Reviewed-by: never, twisti

duke@435	1	/*
twisti@1078	2	* Copyright 1999-2009 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);
cfang@1116	139	bool inline_string_equals();
duke@435	140	Node* pop_math_arg();
duke@435	141	bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
duke@435	142	bool inline_math_native(vmIntrinsics::ID id);
duke@435	143	bool inline_trig(vmIntrinsics::ID id);
duke@435	144	bool inline_trans(vmIntrinsics::ID id);
duke@435	145	bool inline_abs(vmIntrinsics::ID id);
duke@435	146	bool inline_sqrt(vmIntrinsics::ID id);
duke@435	147	bool inline_pow(vmIntrinsics::ID id);
duke@435	148	bool inline_exp(vmIntrinsics::ID id);
duke@435	149	bool inline_min_max(vmIntrinsics::ID id);
duke@435	150	Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
duke@435	151	// This returns Type::AnyPtr, RawPtr, or OopPtr.
duke@435	152	int classify_unsafe_addr(Node* &base, Node* &offset);
duke@435	153	Node* make_unsafe_address(Node* base, Node* offset);
duke@435	154	bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
duke@435	155	bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
duke@435	156	bool inline_unsafe_allocate();
duke@435	157	bool inline_unsafe_copyMemory();
duke@435	158	bool inline_native_currentThread();
duke@435	159	bool inline_native_time_funcs(bool isNano);
duke@435	160	bool inline_native_isInterrupted();
duke@435	161	bool inline_native_Class_query(vmIntrinsics::ID id);
duke@435	162	bool inline_native_subtype_check();
duke@435	163
duke@435	164	bool inline_native_newArray();
duke@435	165	bool inline_native_getLength();
duke@435	166	bool inline_array_copyOf(bool is_copyOfRange);
rasbold@604	167	bool inline_array_equals();
kvn@1268	168	void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
duke@435	169	bool inline_native_clone(bool is_virtual);
duke@435	170	bool inline_native_Reflection_getCallerClass();
duke@435	171	bool inline_native_AtomicLong_get();
duke@435	172	bool inline_native_AtomicLong_attemptUpdate();
duke@435	173	bool is_method_invoke_or_aux_frame(JVMState* jvms);
duke@435	174	// Helper function for inlining native object hash method
duke@435	175	bool inline_native_hashcode(bool is_virtual, bool is_static);
duke@435	176	bool inline_native_getClass();
duke@435	177
duke@435	178	// Helper functions for inlining arraycopy
duke@435	179	bool inline_arraycopy();
duke@435	180	void generate_arraycopy(const TypePtr* adr_type,
duke@435	181	BasicType basic_elem_type,
duke@435	182	Node* src, Node* src_offset,
duke@435	183	Node* dest, Node* dest_offset,
duke@435	184	Node* copy_length,
duke@435	185	bool disjoint_bases = false,
duke@435	186	bool length_never_negative = false,
duke@435	187	RegionNode* slow_region = NULL);
duke@435	188	AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
duke@435	189	RegionNode* slow_region);
duke@435	190	void generate_clear_array(const TypePtr* adr_type,
duke@435	191	Node* dest,
duke@435	192	BasicType basic_elem_type,
duke@435	193	Node* slice_off,
duke@435	194	Node* slice_len,
duke@435	195	Node* slice_end);
duke@435	196	bool generate_block_arraycopy(const TypePtr* adr_type,
duke@435	197	BasicType basic_elem_type,
duke@435	198	AllocateNode* alloc,
duke@435	199	Node* src, Node* src_offset,
duke@435	200	Node* dest, Node* dest_offset,
duke@435	201	Node* dest_size);
duke@435	202	void generate_slow_arraycopy(const TypePtr* adr_type,
duke@435	203	Node* src, Node* src_offset,
duke@435	204	Node* dest, Node* dest_offset,
kvn@1268	205	Node* copy_length);
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,
kvn@1268	210	Node* copy_length);
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,
kvn@1268	214	Node* copy_length);
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);
twisti@1210	224	bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
twisti@1210	225	bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
twisti@1078	226	bool inline_bitCount(vmIntrinsics::ID id);
duke@435	227	bool inline_reverseBytes(vmIntrinsics::ID id);
duke@435	228	};
duke@435	229
duke@435	230
duke@435	231	//---------------------------make_vm_intrinsic----------------------------
duke@435	232	CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
duke@435	233	vmIntrinsics::ID id = m->intrinsic_id();
duke@435	234	assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
duke@435	235
duke@435	236	if (DisableIntrinsic[0] != '\0'
duke@435	237	&& strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
duke@435	238	// disabled by a user request on the command line:
duke@435	239	// example: -XX:DisableIntrinsic=_hashCode,_getClass
duke@435	240	return NULL;
duke@435	241	}
duke@435	242
duke@435	243	if (!m->is_loaded()) {
duke@435	244	// do not attempt to inline unloaded methods
duke@435	245	return NULL;
duke@435	246	}
duke@435	247
duke@435	248	// Only a few intrinsics implement a virtual dispatch.
duke@435	249	// They are expensive calls which are also frequently overridden.
duke@435	250	if (is_virtual) {
duke@435	251	switch (id) {
duke@435	252	case vmIntrinsics::_hashCode:
duke@435	253	case vmIntrinsics::_clone:
duke@435	254	// OK, Object.hashCode and Object.clone intrinsics come in both flavors
duke@435	255	break;
duke@435	256	default:
duke@435	257	return NULL;
duke@435	258	}
duke@435	259	}
duke@435	260
duke@435	261	// -XX:-InlineNatives disables nearly all intrinsics:
duke@435	262	if (!InlineNatives) {
duke@435	263	switch (id) {
duke@435	264	case vmIntrinsics::_indexOf:
duke@435	265	case vmIntrinsics::_compareTo:
cfang@1116	266	case vmIntrinsics::_equals:
rasbold@604	267	case vmIntrinsics::_equalsC:
duke@435	268	break; // InlineNatives does not control String.compareTo
duke@435	269	default:
duke@435	270	return NULL;
duke@435	271	}
duke@435	272	}
duke@435	273
duke@435	274	switch (id) {
duke@435	275	case vmIntrinsics::_compareTo:
duke@435	276	if (!SpecialStringCompareTo) return NULL;
duke@435	277	break;
duke@435	278	case vmIntrinsics::_indexOf:
duke@435	279	if (!SpecialStringIndexOf) return NULL;
duke@435	280	break;
cfang@1116	281	case vmIntrinsics::_equals:
cfang@1116	282	if (!SpecialStringEquals) return NULL;
cfang@1116	283	break;
rasbold@604	284	case vmIntrinsics::_equalsC:
rasbold@604	285	if (!SpecialArraysEquals) return NULL;
rasbold@604	286	break;
duke@435	287	case vmIntrinsics::_arraycopy:
duke@435	288	if (!InlineArrayCopy) return NULL;
duke@435	289	break;
duke@435	290	case vmIntrinsics::_copyMemory:
duke@435	291	if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
duke@435	292	if (!InlineArrayCopy) return NULL;
duke@435	293	break;
duke@435	294	case vmIntrinsics::_hashCode:
duke@435	295	if (!InlineObjectHash) return NULL;
duke@435	296	break;
duke@435	297	case vmIntrinsics::_clone:
duke@435	298	case vmIntrinsics::_copyOf:
duke@435	299	case vmIntrinsics::_copyOfRange:
duke@435	300	if (!InlineObjectCopy) return NULL;
duke@435	301	// These also use the arraycopy intrinsic mechanism:
duke@435	302	if (!InlineArrayCopy) return NULL;
duke@435	303	break;
duke@435	304	case vmIntrinsics::_checkIndex:
duke@435	305	// We do not intrinsify this. The optimizer does fine with it.
duke@435	306	return NULL;
duke@435	307
duke@435	308	case vmIntrinsics::_get_AtomicLong:
duke@435	309	case vmIntrinsics::_attemptUpdate:
duke@435	310	if (!InlineAtomicLong) return NULL;
duke@435	311	break;
duke@435	312
duke@435	313	case vmIntrinsics::_getCallerClass:
duke@435	314	if (!UseNewReflection) return NULL;
duke@435	315	if (!InlineReflectionGetCallerClass) return NULL;
duke@435	316	if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
duke@435	317	break;
duke@435	318
twisti@1078	319	case vmIntrinsics::_bitCount_i:
twisti@1078	320	case vmIntrinsics::_bitCount_l:
twisti@1078	321	if (!UsePopCountInstruction) return NULL;
twisti@1078	322	break;
twisti@1078	323
duke@435	324	default:
jrose@1291	325	assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
jrose@1291	326	assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
duke@435	327	break;
duke@435	328	}
duke@435	329
duke@435	330	// -XX:-InlineClassNatives disables natives from the Class class.
duke@435	331	// The flag applies to all reflective calls, notably Array.newArray
duke@435	332	// (visible to Java programmers as Array.newInstance).
duke@435	333	if (m->holder()->name() == ciSymbol::java_lang_Class() ||
duke@435	334	m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
duke@435	335	if (!InlineClassNatives) return NULL;
duke@435	336	}
duke@435	337
duke@435	338	// -XX:-InlineThreadNatives disables natives from the Thread class.
duke@435	339	if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
duke@435	340	if (!InlineThreadNatives) return NULL;
duke@435	341	}
duke@435	342
duke@435	343	// -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
duke@435	344	if (m->holder()->name() == ciSymbol::java_lang_Math() ||
duke@435	345	m->holder()->name() == ciSymbol::java_lang_Float() ||
duke@435	346	m->holder()->name() == ciSymbol::java_lang_Double()) {
duke@435	347	if (!InlineMathNatives) return NULL;
duke@435	348	}
duke@435	349
duke@435	350	// -XX:-InlineUnsafeOps disables natives from the Unsafe class.
duke@435	351	if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
duke@435	352	if (!InlineUnsafeOps) return NULL;
duke@435	353	}
duke@435	354
duke@435	355	return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
duke@435	356	}
duke@435	357
duke@435	358	//----------------------register_library_intrinsics-----------------------
duke@435	359	// Initialize this file's data structures, for each Compile instance.
duke@435	360	void Compile::register_library_intrinsics() {
duke@435	361	// Nothing to do here.
duke@435	362	}
duke@435	363
duke@435	364	JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
duke@435	365	LibraryCallKit kit(jvms, this);
duke@435	366	Compile* C = kit.C;
duke@435	367	int nodes = C->unique();
duke@435	368	#ifndef PRODUCT
duke@435	369	if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
duke@435	370	char buf[1000];
duke@435	371	const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
duke@435	372	tty->print_cr("Intrinsic %s", str);
duke@435	373	}
duke@435	374	#endif
duke@435	375	if (kit.try_to_inline()) {
duke@435	376	if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
duke@435	377	tty->print("Inlining intrinsic %s%s at bci:%d in",
duke@435	378	vmIntrinsics::name_at(intrinsic_id()),
duke@435	379	(is_virtual() ? " (virtual)" : ""), kit.bci());
duke@435	380	kit.caller()->print_short_name(tty);
duke@435	381	tty->print_cr(" (%d bytes)", kit.caller()->code_size());
duke@435	382	}
duke@435	383	C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
duke@435	384	if (C->log()) {
duke@435	385	C->log()->elem("intrinsic id='%s'%s nodes='%d'",
duke@435	386	vmIntrinsics::name_at(intrinsic_id()),
duke@435	387	(is_virtual() ? " virtual='1'" : ""),
duke@435	388	C->unique() - nodes);
duke@435	389	}
duke@435	390	return kit.transfer_exceptions_into_jvms();
duke@435	391	}
duke@435	392
duke@435	393	if (PrintIntrinsics) {
jrose@1291	394	tty->print("Did not inline intrinsic %s%s at bci:%d in",
jrose@1291	395	vmIntrinsics::name_at(intrinsic_id()),
jrose@1291	396	(is_virtual() ? " (virtual)" : ""), kit.bci());
jrose@1291	397	kit.caller()->print_short_name(tty);
jrose@1291	398	tty->print_cr(" (%d bytes)", kit.caller()->code_size());
duke@435	399	}
duke@435	400	C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
duke@435	401	return NULL;
duke@435	402	}
duke@435	403
duke@435	404	bool LibraryCallKit::try_to_inline() {
duke@435	405	// Handle symbolic names for otherwise undistinguished boolean switches:
duke@435	406	const bool is_store = true;
duke@435	407	const bool is_native_ptr = true;
duke@435	408	const bool is_static = true;
duke@435	409
duke@435	410	switch (intrinsic_id()) {
duke@435	411	case vmIntrinsics::_hashCode:
duke@435	412	return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
duke@435	413	case vmIntrinsics::_identityHashCode:
duke@435	414	return inline_native_hashcode(/*!virtual*/ false, is_static);
duke@435	415	case vmIntrinsics::_getClass:
duke@435	416	return inline_native_getClass();
duke@435	417
duke@435	418	case vmIntrinsics::_dsin:
duke@435	419	case vmIntrinsics::_dcos:
duke@435	420	case vmIntrinsics::_dtan:
duke@435	421	case vmIntrinsics::_dabs:
duke@435	422	case vmIntrinsics::_datan2:
duke@435	423	case vmIntrinsics::_dsqrt:
duke@435	424	case vmIntrinsics::_dexp:
duke@435	425	case vmIntrinsics::_dlog:
duke@435	426	case vmIntrinsics::_dlog10:
duke@435	427	case vmIntrinsics::_dpow:
duke@435	428	return inline_math_native(intrinsic_id());
duke@435	429
duke@435	430	case vmIntrinsics::_min:
duke@435	431	case vmIntrinsics::_max:
duke@435	432	return inline_min_max(intrinsic_id());
duke@435	433
duke@435	434	case vmIntrinsics::_arraycopy:
duke@435	435	return inline_arraycopy();
duke@435	436
duke@435	437	case vmIntrinsics::_compareTo:
duke@435	438	return inline_string_compareTo();
duke@435	439	case vmIntrinsics::_indexOf:
duke@435	440	return inline_string_indexOf();
cfang@1116	441	case vmIntrinsics::_equals:
cfang@1116	442	return inline_string_equals();
duke@435	443
duke@435	444	case vmIntrinsics::_getObject:
duke@435	445	return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
duke@435	446	case vmIntrinsics::_getBoolean:
duke@435	447	return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
duke@435	448	case vmIntrinsics::_getByte:
duke@435	449	return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
duke@435	450	case vmIntrinsics::_getShort:
duke@435	451	return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
duke@435	452	case vmIntrinsics::_getChar:
duke@435	453	return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
duke@435	454	case vmIntrinsics::_getInt:
duke@435	455	return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
duke@435	456	case vmIntrinsics::_getLong:
duke@435	457	return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
duke@435	458	case vmIntrinsics::_getFloat:
duke@435	459	return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
duke@435	460	case vmIntrinsics::_getDouble:
duke@435	461	return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
duke@435	462
duke@435	463	case vmIntrinsics::_putObject:
duke@435	464	return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
duke@435	465	case vmIntrinsics::_putBoolean:
duke@435	466	return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
duke@435	467	case vmIntrinsics::_putByte:
duke@435	468	return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
duke@435	469	case vmIntrinsics::_putShort:
duke@435	470	return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
duke@435	471	case vmIntrinsics::_putChar:
duke@435	472	return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
duke@435	473	case vmIntrinsics::_putInt:
duke@435	474	return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
duke@435	475	case vmIntrinsics::_putLong:
duke@435	476	return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
duke@435	477	case vmIntrinsics::_putFloat:
duke@435	478	return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
duke@435	479	case vmIntrinsics::_putDouble:
duke@435	480	return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
duke@435	481
duke@435	482	case vmIntrinsics::_getByte_raw:
duke@435	483	return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
duke@435	484	case vmIntrinsics::_getShort_raw:
duke@435	485	return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
duke@435	486	case vmIntrinsics::_getChar_raw:
duke@435	487	return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
duke@435	488	case vmIntrinsics::_getInt_raw:
duke@435	489	return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
duke@435	490	case vmIntrinsics::_getLong_raw:
duke@435	491	return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
duke@435	492	case vmIntrinsics::_getFloat_raw:
duke@435	493	return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
duke@435	494	case vmIntrinsics::_getDouble_raw:
duke@435	495	return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
duke@435	496	case vmIntrinsics::_getAddress_raw:
duke@435	497	return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
duke@435	498
duke@435	499	case vmIntrinsics::_putByte_raw:
duke@435	500	return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
duke@435	501	case vmIntrinsics::_putShort_raw:
duke@435	502	return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
duke@435	503	case vmIntrinsics::_putChar_raw:
duke@435	504	return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
duke@435	505	case vmIntrinsics::_putInt_raw:
duke@435	506	return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
duke@435	507	case vmIntrinsics::_putLong_raw:
duke@435	508	return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
duke@435	509	case vmIntrinsics::_putFloat_raw:
duke@435	510	return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
duke@435	511	case vmIntrinsics::_putDouble_raw:
duke@435	512	return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
duke@435	513	case vmIntrinsics::_putAddress_raw:
duke@435	514	return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
duke@435	515
duke@435	516	case vmIntrinsics::_getObjectVolatile:
duke@435	517	return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
duke@435	518	case vmIntrinsics::_getBooleanVolatile:
duke@435	519	return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
duke@435	520	case vmIntrinsics::_getByteVolatile:
duke@435	521	return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
duke@435	522	case vmIntrinsics::_getShortVolatile:
duke@435	523	return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
duke@435	524	case vmIntrinsics::_getCharVolatile:
duke@435	525	return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
duke@435	526	case vmIntrinsics::_getIntVolatile:
duke@435	527	return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
duke@435	528	case vmIntrinsics::_getLongVolatile:
duke@435	529	return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
duke@435	530	case vmIntrinsics::_getFloatVolatile:
duke@435	531	return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
duke@435	532	case vmIntrinsics::_getDoubleVolatile:
duke@435	533	return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
duke@435	534
duke@435	535	case vmIntrinsics::_putObjectVolatile:
duke@435	536	return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
duke@435	537	case vmIntrinsics::_putBooleanVolatile:
duke@435	538	return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
duke@435	539	case vmIntrinsics::_putByteVolatile:
duke@435	540	return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
duke@435	541	case vmIntrinsics::_putShortVolatile:
duke@435	542	return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
duke@435	543	case vmIntrinsics::_putCharVolatile:
duke@435	544	return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
duke@435	545	case vmIntrinsics::_putIntVolatile:
duke@435	546	return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
duke@435	547	case vmIntrinsics::_putLongVolatile:
duke@435	548	return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
duke@435	549	case vmIntrinsics::_putFloatVolatile:
duke@435	550	return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
duke@435	551	case vmIntrinsics::_putDoubleVolatile:
duke@435	552	return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
duke@435	553
duke@435	554	case vmIntrinsics::_prefetchRead:
duke@435	555	return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
duke@435	556	case vmIntrinsics::_prefetchWrite:
duke@435	557	return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
duke@435	558	case vmIntrinsics::_prefetchReadStatic:
duke@435	559	return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
duke@435	560	case vmIntrinsics::_prefetchWriteStatic:
duke@435	561	return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
duke@435	562
duke@435	563	case vmIntrinsics::_compareAndSwapObject:
duke@435	564	return inline_unsafe_CAS(T_OBJECT);
duke@435	565	case vmIntrinsics::_compareAndSwapInt:
duke@435	566	return inline_unsafe_CAS(T_INT);
duke@435	567	case vmIntrinsics::_compareAndSwapLong:
duke@435	568	return inline_unsafe_CAS(T_LONG);
duke@435	569
duke@435	570	case vmIntrinsics::_putOrderedObject:
duke@435	571	return inline_unsafe_ordered_store(T_OBJECT);
duke@435	572	case vmIntrinsics::_putOrderedInt:
duke@435	573	return inline_unsafe_ordered_store(T_INT);
duke@435	574	case vmIntrinsics::_putOrderedLong:
duke@435	575	return inline_unsafe_ordered_store(T_LONG);
duke@435	576
duke@435	577	case vmIntrinsics::_currentThread:
duke@435	578	return inline_native_currentThread();
duke@435	579	case vmIntrinsics::_isInterrupted:
duke@435	580	return inline_native_isInterrupted();
duke@435	581
duke@435	582	case vmIntrinsics::_currentTimeMillis:
duke@435	583	return inline_native_time_funcs(false);
duke@435	584	case vmIntrinsics::_nanoTime:
duke@435	585	return inline_native_time_funcs(true);
duke@435	586	case vmIntrinsics::_allocateInstance:
duke@435	587	return inline_unsafe_allocate();
duke@435	588	case vmIntrinsics::_copyMemory:
duke@435	589	return inline_unsafe_copyMemory();
duke@435	590	case vmIntrinsics::_newArray:
duke@435	591	return inline_native_newArray();
duke@435	592	case vmIntrinsics::_getLength:
duke@435	593	return inline_native_getLength();
duke@435	594	case vmIntrinsics::_copyOf:
duke@435	595	return inline_array_copyOf(false);
duke@435	596	case vmIntrinsics::_copyOfRange:
duke@435	597	return inline_array_copyOf(true);
rasbold@604	598	case vmIntrinsics::_equalsC:
rasbold@604	599	return inline_array_equals();
duke@435	600	case vmIntrinsics::_clone:
duke@435	601	return inline_native_clone(intrinsic()->is_virtual());
duke@435	602
duke@435	603	case vmIntrinsics::_isAssignableFrom:
duke@435	604	return inline_native_subtype_check();
duke@435	605
duke@435	606	case vmIntrinsics::_isInstance:
duke@435	607	case vmIntrinsics::_getModifiers:
duke@435	608	case vmIntrinsics::_isInterface:
duke@435	609	case vmIntrinsics::_isArray:
duke@435	610	case vmIntrinsics::_isPrimitive:
duke@435	611	case vmIntrinsics::_getSuperclass:
duke@435	612	case vmIntrinsics::_getComponentType:
duke@435	613	case vmIntrinsics::_getClassAccessFlags:
duke@435	614	return inline_native_Class_query(intrinsic_id());
duke@435	615
duke@435	616	case vmIntrinsics::_floatToRawIntBits:
duke@435	617	case vmIntrinsics::_floatToIntBits:
duke@435	618	case vmIntrinsics::_intBitsToFloat:
duke@435	619	case vmIntrinsics::_doubleToRawLongBits:
duke@435	620	case vmIntrinsics::_doubleToLongBits:
duke@435	621	case vmIntrinsics::_longBitsToDouble:
duke@435	622	return inline_fp_conversions(intrinsic_id());
duke@435	623
twisti@1210	624	case vmIntrinsics::_numberOfLeadingZeros_i:
twisti@1210	625	case vmIntrinsics::_numberOfLeadingZeros_l:
twisti@1210	626	return inline_numberOfLeadingZeros(intrinsic_id());
twisti@1210	627
twisti@1210	628	case vmIntrinsics::_numberOfTrailingZeros_i:
twisti@1210	629	case vmIntrinsics::_numberOfTrailingZeros_l:
twisti@1210	630	return inline_numberOfTrailingZeros(intrinsic_id());
twisti@1210	631
twisti@1078	632	case vmIntrinsics::_bitCount_i:
twisti@1078	633	case vmIntrinsics::_bitCount_l:
twisti@1078	634	return inline_bitCount(intrinsic_id());
twisti@1078	635
duke@435	636	case vmIntrinsics::_reverseBytes_i:
duke@435	637	case vmIntrinsics::_reverseBytes_l:
duke@435	638	return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
duke@435	639
duke@435	640	case vmIntrinsics::_get_AtomicLong:
duke@435	641	return inline_native_AtomicLong_get();
duke@435	642	case vmIntrinsics::_attemptUpdate:
duke@435	643	return inline_native_AtomicLong_attemptUpdate();
duke@435	644
duke@435	645	case vmIntrinsics::_getCallerClass:
duke@435	646	return inline_native_Reflection_getCallerClass();
duke@435	647
duke@435	648	default:
duke@435	649	// If you get here, it may be that someone has added a new intrinsic
duke@435	650	// to the list in vmSymbols.hpp without implementing it here.
duke@435	651	#ifndef PRODUCT
duke@435	652	if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
duke@435	653	tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
duke@435	654	vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
duke@435	655	}
duke@435	656	#endif
duke@435	657	return false;
duke@435	658	}
duke@435	659	}
duke@435	660
duke@435	661	//------------------------------push_result------------------------------
duke@435	662	// Helper function for finishing intrinsics.
duke@435	663	void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
duke@435	664	record_for_igvn(region);
duke@435	665	set_control(_gvn.transform(region));
duke@435	666	BasicType value_type = value->type()->basic_type();
duke@435	667	push_node(value_type, _gvn.transform(value));
duke@435	668	}
duke@435	669
duke@435	670	//------------------------------generate_guard---------------------------
duke@435	671	// Helper function for generating guarded fast-slow graph structures.
duke@435	672	// The given 'test', if true, guards a slow path. If the test fails
duke@435	673	// then a fast path can be taken. (We generally hope it fails.)
duke@435	674	// In all cases, GraphKit::control() is updated to the fast path.
duke@435	675	// The returned value represents the control for the slow path.
duke@435	676	// The return value is never 'top'; it is either a valid control
duke@435	677	// or NULL if it is obvious that the slow path can never be taken.
duke@435	678	// Also, if region and the slow control are not NULL, the slow edge
duke@435	679	// is appended to the region.
duke@435	680	Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
duke@435	681	if (stopped()) {
duke@435	682	// Already short circuited.
duke@435	683	return NULL;
duke@435	684	}
duke@435	685
duke@435	686	// Build an if node and its projections.
duke@435	687	// If test is true we take the slow path, which we assume is uncommon.
duke@435	688	if (_gvn.type(test) == TypeInt::ZERO) {
duke@435	689	// The slow branch is never taken. No need to build this guard.
duke@435	690	return NULL;
duke@435	691	}
duke@435	692
duke@435	693	IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
duke@435	694
duke@435	695	Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
duke@435	696	if (if_slow == top()) {
duke@435	697	// The slow branch is never taken. No need to build this guard.
duke@435	698	return NULL;
duke@435	699	}
duke@435	700
duke@435	701	if (region != NULL)
duke@435	702	region->add_req(if_slow);
duke@435	703
duke@435	704	Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
duke@435	705	set_control(if_fast);
duke@435	706
duke@435	707	return if_slow;
duke@435	708	}
duke@435	709
duke@435	710	inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
duke@435	711	return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
duke@435	712	}
duke@435	713	inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
duke@435	714	return generate_guard(test, region, PROB_FAIR);
duke@435	715	}
duke@435	716
duke@435	717	inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
duke@435	718	Node* *pos_index) {
duke@435	719	if (stopped())
duke@435	720	return NULL; // already stopped
duke@435	721	if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
duke@435	722	return NULL; // index is already adequately typed
duke@435	723	Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
duke@435	724	Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
duke@435	725	Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
duke@435	726	if (is_neg != NULL && pos_index != NULL) {
duke@435	727	// Emulate effect of Parse::adjust_map_after_if.
duke@435	728	Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
duke@435	729	ccast->set_req(0, control());
duke@435	730	(*pos_index) = _gvn.transform(ccast);
duke@435	731	}
duke@435	732	return is_neg;
duke@435	733	}
duke@435	734
duke@435	735	inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
duke@435	736	Node* *pos_index) {
duke@435	737	if (stopped())
duke@435	738	return NULL; // already stopped
duke@435	739	if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
duke@435	740	return NULL; // index is already adequately typed
duke@435	741	Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
duke@435	742	BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
duke@435	743	Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
duke@435	744	Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
duke@435	745	if (is_notp != NULL && pos_index != NULL) {
duke@435	746	// Emulate effect of Parse::adjust_map_after_if.
duke@435	747	Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
duke@435	748	ccast->set_req(0, control());
duke@435	749	(*pos_index) = _gvn.transform(ccast);
duke@435	750	}
duke@435	751	return is_notp;
duke@435	752	}
duke@435	753
duke@435	754	// Make sure that 'position' is a valid limit index, in [0..length].
duke@435	755	// There are two equivalent plans for checking this:
duke@435	756	// A. (offset + copyLength) unsigned<= arrayLength
duke@435	757	// B. offset <= (arrayLength - copyLength)
duke@435	758	// We require that all of the values above, except for the sum and
duke@435	759	// difference, are already known to be non-negative.
duke@435	760	// Plan A is robust in the face of overflow, if offset and copyLength
duke@435	761	// are both hugely positive.
duke@435	762	//
duke@435	763	// Plan B is less direct and intuitive, but it does not overflow at
duke@435	764	// all, since the difference of two non-negatives is always
duke@435	765	// representable. Whenever Java methods must perform the equivalent
duke@435	766	// check they generally use Plan B instead of Plan A.
duke@435	767	// For the moment we use Plan A.
duke@435	768	inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
duke@435	769	Node* subseq_length,
duke@435	770	Node* array_length,
duke@435	771	RegionNode* region) {
duke@435	772	if (stopped())
duke@435	773	return NULL; // already stopped
duke@435	774	bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
duke@435	775	if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
duke@435	776	return NULL; // common case of whole-array copy
duke@435	777	Node* last = subseq_length;
duke@435	778	if (!zero_offset) // last += offset
duke@435	779	last = _gvn.transform( new (C, 3) AddINode(last, offset));
duke@435	780	Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
duke@435	781	Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
duke@435	782	Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
duke@435	783	return is_over;
duke@435	784	}
duke@435	785
duke@435	786
duke@435	787	//--------------------------generate_current_thread--------------------
duke@435	788	Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
duke@435	789	ciKlass* thread_klass = env()->Thread_klass();
duke@435	790	const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
duke@435	791	Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
duke@435	792	Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
duke@435	793	Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
duke@435	794	tls_output = thread;
duke@435	795	return threadObj;
duke@435	796	}
duke@435	797
duke@435	798
duke@435	799	//------------------------------inline_string_compareTo------------------------
duke@435	800	bool LibraryCallKit::inline_string_compareTo() {
duke@435	801
cfang@1116	802	if (!Matcher::has_match_rule(Op_StrComp)) return false;
cfang@1116	803
duke@435	804	const int value_offset = java_lang_String::value_offset_in_bytes();
duke@435	805	const int count_offset = java_lang_String::count_offset_in_bytes();
duke@435	806	const int offset_offset = java_lang_String::offset_offset_in_bytes();
duke@435	807
duke@435	808	_sp += 2;
duke@435	809	Node *argument = pop(); // pop non-receiver first: it was pushed second
duke@435	810	Node *receiver = pop();
duke@435	811
duke@435	812	// Null check on self without removing any arguments. The argument
duke@435	813	// null check technically happens in the wrong place, which can lead to
duke@435	814	// invalid stack traces when string compare is inlined into a method
duke@435	815	// which handles NullPointerExceptions.
duke@435	816	_sp += 2;
duke@435	817	receiver = do_null_check(receiver, T_OBJECT);
duke@435	818	argument = do_null_check(argument, T_OBJECT);
duke@435	819	_sp -= 2;
duke@435	820	if (stopped()) {
duke@435	821	return true;
duke@435	822	}
duke@435	823
duke@435	824	ciInstanceKlass* klass = env()->String_klass();
duke@435	825	const TypeInstPtr* string_type =
duke@435	826	TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
duke@435	827
duke@435	828	Node* compare =
duke@435	829	_gvn.transform(new (C, 7) StrCompNode(
duke@435	830	control(),
duke@435	831	memory(TypeAryPtr::CHARS),
duke@435	832	memory(string_type->add_offset(value_offset)),
duke@435	833	memory(string_type->add_offset(count_offset)),
duke@435	834	memory(string_type->add_offset(offset_offset)),
duke@435	835	receiver,
duke@435	836	argument));
duke@435	837	push(compare);
duke@435	838	return true;
duke@435	839	}
duke@435	840
cfang@1116	841	//------------------------------inline_string_equals------------------------
cfang@1116	842	bool LibraryCallKit::inline_string_equals() {
cfang@1116	843
cfang@1116	844	if (!Matcher::has_match_rule(Op_StrEquals)) return false;
cfang@1116	845
cfang@1116	846	const int value_offset = java_lang_String::value_offset_in_bytes();
cfang@1116	847	const int count_offset = java_lang_String::count_offset_in_bytes();
cfang@1116	848	const int offset_offset = java_lang_String::offset_offset_in_bytes();
cfang@1116	849
cfang@1116	850	_sp += 2;
cfang@1116	851	Node* argument = pop(); // pop non-receiver first: it was pushed second
cfang@1116	852	Node* receiver = pop();
cfang@1116	853
cfang@1116	854	// Null check on self without removing any arguments. The argument
cfang@1116	855	// null check technically happens in the wrong place, which can lead to
cfang@1116	856	// invalid stack traces when string compare is inlined into a method
cfang@1116	857	// which handles NullPointerExceptions.
cfang@1116	858	_sp += 2;
cfang@1116	859	receiver = do_null_check(receiver, T_OBJECT);
cfang@1116	860	//should not do null check for argument for String.equals(), because spec
cfang@1116	861	//allows to specify NULL as argument.
cfang@1116	862	_sp -= 2;
cfang@1116	863
cfang@1116	864	if (stopped()) {
cfang@1116	865	return true;
cfang@1116	866	}
cfang@1116	867
cfang@1116	868	// get String klass for instanceOf
cfang@1116	869	ciInstanceKlass* klass = env()->String_klass();
cfang@1116	870
cfang@1116	871	// two paths (plus control) merge
cfang@1116	872	RegionNode* region = new (C, 3) RegionNode(3);
cfang@1116	873	Node* phi = new (C, 3) PhiNode(region, TypeInt::BOOL);
cfang@1116	874
cfang@1116	875	Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
cfang@1116	876	Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
cfang@1116	877	Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
cfang@1116	878
cfang@1116	879	IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
cfang@1116	880
cfang@1116	881	Node* if_true = _gvn.transform(new (C, 1) IfTrueNode(iff));
cfang@1116	882	set_control(if_true);
cfang@1116	883
cfang@1116	884	const TypeInstPtr* string_type =
cfang@1116	885	TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
cfang@1116	886
cfang@1116	887	// instanceOf == true
cfang@1116	888	Node* equals =
cfang@1116	889	_gvn.transform(new (C, 7) StrEqualsNode(
cfang@1116	890	control(),
cfang@1116	891	memory(TypeAryPtr::CHARS),
cfang@1116	892	memory(string_type->add_offset(value_offset)),
cfang@1116	893	memory(string_type->add_offset(count_offset)),
cfang@1116	894	memory(string_type->add_offset(offset_offset)),
cfang@1116	895	receiver,
cfang@1116	896	argument));
cfang@1116	897
cfang@1116	898	phi->init_req(1, _gvn.transform(equals));
cfang@1116	899	region->init_req(1, if_true);
cfang@1116	900
cfang@1116	901	//instanceOf == false, fallthrough
cfang@1116	902	Node* if_false = _gvn.transform(new (C, 1) IfFalseNode(iff));
cfang@1116	903	set_control(if_false);
cfang@1116	904
cfang@1116	905	phi->init_req(2, _gvn.transform(intcon(0)));
cfang@1116	906	region->init_req(2, if_false);
cfang@1116	907
cfang@1116	908	// post merge
cfang@1116	909	set_control(_gvn.transform(region));
cfang@1116	910	record_for_igvn(region);
cfang@1116	911
cfang@1116	912	push(_gvn.transform(phi));
cfang@1116	913
cfang@1116	914	return true;
cfang@1116	915	}
cfang@1116	916
rasbold@604	917	//------------------------------inline_array_equals----------------------------
rasbold@604	918	bool LibraryCallKit::inline_array_equals() {
rasbold@604	919
rasbold@609	920	if (!Matcher::has_match_rule(Op_AryEq)) return false;
rasbold@609	921
rasbold@604	922	_sp += 2;
rasbold@604	923	Node *argument2 = pop();
rasbold@604	924	Node *argument1 = pop();
rasbold@604	925
rasbold@604	926	Node* equals =
rasbold@604	927	_gvn.transform(new (C, 3) AryEqNode(control(),
rasbold@604	928	argument1,
rasbold@604	929	argument2)
rasbold@604	930	);
rasbold@604	931	push(equals);
rasbold@604	932	return true;
rasbold@604	933	}
rasbold@604	934
duke@435	935	// Java version of String.indexOf(constant string)
duke@435	936	// class StringDecl {
duke@435	937	// StringDecl(char[] ca) {
duke@435	938	// offset = 0;
duke@435	939	// count = ca.length;
duke@435	940	// value = ca;
duke@435	941	// }
duke@435	942	// int offset;
duke@435	943	// int count;
duke@435	944	// char[] value;
duke@435	945	// }
duke@435	946	//
duke@435	947	// static int string_indexOf_J(StringDecl string_object, char[] target_object,
duke@435	948	// int targetOffset, int cache_i, int md2) {
duke@435	949	// int cache = cache_i;
duke@435	950	// int sourceOffset = string_object.offset;
duke@435	951	// int sourceCount = string_object.count;
duke@435	952	// int targetCount = target_object.length;
duke@435	953	//
duke@435	954	// int targetCountLess1 = targetCount - 1;
duke@435	955	// int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
duke@435	956	//
duke@435	957	// char[] source = string_object.value;
duke@435	958	// char[] target = target_object;
duke@435	959	// int lastChar = target[targetCountLess1];
duke@435	960	//
duke@435	961	// outer_loop:
duke@435	962	// for (int i = sourceOffset; i < sourceEnd; ) {
duke@435	963	// int src = source[i + targetCountLess1];
duke@435	964	// if (src == lastChar) {
duke@435	965	// // With random strings and a 4-character alphabet,
duke@435	966	// // reverse matching at this point sets up 0.8% fewer
duke@435	967	// // frames, but (paradoxically) makes 0.3% more probes.
duke@435	968	// // Since those probes are nearer the lastChar probe,
duke@435	969	// // there is may be a net D$ win with reverse matching.
duke@435	970	// // But, reversing loop inhibits unroll of inner loop
duke@435	971	// // for unknown reason. So, does running outer loop from
duke@435	972	// // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
duke@435	973	// for (int j = 0; j < targetCountLess1; j++) {
duke@435	974	// if (target[targetOffset + j] != source[i+j]) {
duke@435	975	// if ((cache & (1 << source[i+j])) == 0) {
duke@435	976	// if (md2 < j+1) {
duke@435	977	// i += j+1;
duke@435	978	// continue outer_loop;
duke@435	979	// }
duke@435	980	// }
duke@435	981	// i += md2;
duke@435	982	// continue outer_loop;
duke@435	983	// }
duke@435	984	// }
duke@435	985	// return i - sourceOffset;
duke@435	986	// }
duke@435	987	// if ((cache & (1 << src)) == 0) {
duke@435	988	// i += targetCountLess1;
duke@435	989	// } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
duke@435	990	// i++;
duke@435	991	// }
duke@435	992	// return -1;
duke@435	993	// }
duke@435	994
duke@435	995	//------------------------------string_indexOf------------------------
duke@435	996	Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
duke@435	997	jint cache_i, jint md2_i) {
duke@435	998
duke@435	999	Node* no_ctrl = NULL;
duke@435	1000	float likely = PROB_LIKELY(0.9);
duke@435	1001	float unlikely = PROB_UNLIKELY(0.9);
duke@435	1002
duke@435	1003	const int value_offset = java_lang_String::value_offset_in_bytes();
duke@435	1004	const int count_offset = java_lang_String::count_offset_in_bytes();
duke@435	1005	const int offset_offset = java_lang_String::offset_offset_in_bytes();
duke@435	1006
duke@435	1007	ciInstanceKlass* klass = env()->String_klass();
duke@435	1008	const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
duke@435	1009	const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
duke@435	1010
duke@435	1011	Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
duke@435	1012	Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
duke@435	1013	Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
duke@435	1014	Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
duke@435	1015	Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
duke@435	1016	Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
duke@435	1017
kvn@599	1018	Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) );
duke@435	1019	jint target_length = target_array->length();
duke@435	1020	const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
duke@435	1021	const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
duke@435	1022
kvn@1286	1023	IdealKit kit(gvn(), control(), merged_memory(), false, true);
duke@435	1024	#define __ kit.
duke@435	1025	Node* zero = __ ConI(0);
duke@435	1026	Node* one = __ ConI(1);
duke@435	1027	Node* cache = __ ConI(cache_i);
duke@435	1028	Node* md2 = __ ConI(md2_i);
duke@435	1029	Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
duke@435	1030	Node* targetCount = __ ConI(target_length);
duke@435	1031	Node* targetCountLess1 = __ ConI(target_length - 1);
duke@435	1032	Node* targetOffset = __ ConI(targetOffset_i);
duke@435	1033	Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
duke@435	1034
kvn@1286	1035	IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
duke@435	1036	Node* outer_loop = __ make_label(2 /* goto */);
duke@435	1037	Node* return_ = __ make_label(1);
duke@435	1038
duke@435	1039	__ set(rtn,__ ConI(-1));
duke@435	1040	__ loop(i, sourceOffset, BoolTest::lt, sourceEnd); {
duke@435	1041	Node* i2 = __ AddI(__ value(i), targetCountLess1);
duke@435	1042	// pin to prohibit loading of "next iteration" value which may SEGV (rare)
duke@435	1043	Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
duke@435	1044	__ if_then(src, BoolTest::eq, lastChar, unlikely); {
duke@435	1045	__ loop(j, zero, BoolTest::lt, targetCountLess1); {
duke@435	1046	Node* tpj = __ AddI(targetOffset, __ value(j));
duke@435	1047	Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
duke@435	1048	Node* ipj = __ AddI(__ value(i), __ value(j));
duke@435	1049	Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
duke@435	1050	__ if_then(targ, BoolTest::ne, src2); {
duke@435	1051	__ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
duke@435	1052	__ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
duke@435	1053	__ increment(i, __ AddI(__ value(j), one));
duke@435	1054	__ goto_(outer_loop);
duke@435	1055	} __ end_if(); __ dead(j);
duke@435	1056	}__ end_if(); __ dead(j);
duke@435	1057	__ increment(i, md2);
duke@435	1058	__ goto_(outer_loop);
duke@435	1059	}__ end_if();
duke@435	1060	__ increment(j, one);
duke@435	1061	}__ end_loop(); __ dead(j);
duke@435	1062	__ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
duke@435	1063	__ goto_(return_);
duke@435	1064	}__ end_if();
duke@435	1065	__ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
duke@435	1066	__ increment(i, targetCountLess1);
duke@435	1067	}__ end_if();
duke@435	1068	__ increment(i, one);
duke@435	1069	__ bind(outer_loop);
duke@435	1070	}__ end_loop(); __ dead(i);
duke@435	1071	__ bind(return_);
kvn@1286	1072
kvn@1286	1073	// Final sync IdealKit and GraphKit.
kvn@1286	1074	sync_kit(kit);
duke@435	1075	Node* result = __ value(rtn);
duke@435	1076	#undef __
duke@435	1077	C->set_has_loops(true);
duke@435	1078	return result;
duke@435	1079	}
duke@435	1080
duke@435	1081	//------------------------------inline_string_indexOf------------------------
duke@435	1082	bool LibraryCallKit::inline_string_indexOf() {
duke@435	1083
cfang@1116	1084	const int value_offset = java_lang_String::value_offset_in_bytes();
cfang@1116	1085	const int count_offset = java_lang_String::count_offset_in_bytes();
cfang@1116	1086	const int offset_offset = java_lang_String::offset_offset_in_bytes();
cfang@1116	1087
duke@435	1088	_sp += 2;
duke@435	1089	Node *argument = pop(); // pop non-receiver first: it was pushed second
duke@435	1090	Node *receiver = pop();
duke@435	1091
cfang@1116	1092	Node* result;
cfang@1116	1093	if (Matcher::has_match_rule(Op_StrIndexOf) &&
cfang@1116	1094	UseSSE42Intrinsics) {
cfang@1116	1095	// Generate SSE4.2 version of indexOf
cfang@1116	1096	// We currently only have match rules that use SSE4.2
cfang@1116	1097
cfang@1116	1098	// Null check on self without removing any arguments. The argument
cfang@1116	1099	// null check technically happens in the wrong place, which can lead to
cfang@1116	1100	// invalid stack traces when string compare is inlined into a method
cfang@1116	1101	// which handles NullPointerExceptions.
cfang@1116	1102	_sp += 2;
cfang@1116	1103	receiver = do_null_check(receiver, T_OBJECT);
cfang@1116	1104	argument = do_null_check(argument, T_OBJECT);
cfang@1116	1105	_sp -= 2;
cfang@1116	1106
cfang@1116	1107	if (stopped()) {
cfang@1116	1108	return true;
cfang@1116	1109	}
cfang@1116	1110
cfang@1116	1111	ciInstanceKlass* klass = env()->String_klass();
cfang@1116	1112	const TypeInstPtr* string_type =
cfang@1116	1113	TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0);
cfang@1116	1114
cfang@1116	1115	result =
cfang@1116	1116	_gvn.transform(new (C, 7)
cfang@1116	1117	StrIndexOfNode(control(),
cfang@1116	1118	memory(TypeAryPtr::CHARS),
cfang@1116	1119	memory(string_type->add_offset(value_offset)),
cfang@1116	1120	memory(string_type->add_offset(count_offset)),
cfang@1116	1121	memory(string_type->add_offset(offset_offset)),
cfang@1116	1122	receiver,
cfang@1116	1123	argument));
cfang@1116	1124	} else { //Use LibraryCallKit::string_indexOf
cfang@1116	1125	// don't intrinsify is argument isn't a constant string.
cfang@1116	1126	if (!argument->is_Con()) {
cfang@1116	1127	return false;
cfang@1116	1128	}
cfang@1116	1129	const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
cfang@1116	1130	if (str_type == NULL) {
cfang@1116	1131	return false;
cfang@1116	1132	}
cfang@1116	1133	ciInstanceKlass* klass = env()->String_klass();
cfang@1116	1134	ciObject* str_const = str_type->const_oop();
cfang@1116	1135	if (str_const == NULL || str_const->klass() != klass) {
cfang@1116	1136	return false;
cfang@1116	1137	}
cfang@1116	1138	ciInstance* str = str_const->as_instance();
cfang@1116	1139	assert(str != NULL, "must be instance");
cfang@1116	1140
cfang@1116	1141	ciObject* v = str->field_value_by_offset(value_offset).as_object();
cfang@1116	1142	int o = str->field_value_by_offset(offset_offset).as_int();
cfang@1116	1143	int c = str->field_value_by_offset(count_offset).as_int();
cfang@1116	1144	ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
cfang@1116	1145
cfang@1116	1146	// constant strings have no offset and count == length which
cfang@1116	1147	// simplifies the resulting code somewhat so lets optimize for that.
cfang@1116	1148	if (o != 0 || c != pat->length()) {
cfang@1116	1149	return false;
cfang@1116	1150	}
cfang@1116	1151
cfang@1116	1152	// Null check on self without removing any arguments. The argument
cfang@1116	1153	// null check technically happens in the wrong place, which can lead to
cfang@1116	1154	// invalid stack traces when string compare is inlined into a method
cfang@1116	1155	// which handles NullPointerExceptions.
cfang@1116	1156	_sp += 2;
cfang@1116	1157	receiver = do_null_check(receiver, T_OBJECT);
cfang@1116	1158	// No null check on the argument is needed since it's a constant String oop.
cfang@1116	1159	_sp -= 2;
cfang@1116	1160	if (stopped()) {
cfang@1116	1161	return true;
cfang@1116	1162	}
cfang@1116	1163
cfang@1116	1164	// The null string as a pattern always returns 0 (match at beginning of string)
cfang@1116	1165	if (c == 0) {
cfang@1116	1166	push(intcon(0));
cfang@1116	1167	return true;
cfang@1116	1168	}
cfang@1116	1169
cfang@1116	1170	// Generate default indexOf
cfang@1116	1171	jchar lastChar = pat->char_at(o + (c - 1));
cfang@1116	1172	int cache = 0;
cfang@1116	1173	int i;
cfang@1116	1174	for (i = 0; i < c - 1; i++) {
cfang@1116	1175	assert(i < pat->length(), "out of range");
cfang@1116	1176	cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
cfang@1116	1177	}
cfang@1116	1178
cfang@1116	1179	int md2 = c;
cfang@1116	1180	for (i = 0; i < c - 1; i++) {
cfang@1116	1181	assert(i < pat->length(), "out of range");
cfang@1116	1182	if (pat->char_at(o + i) == lastChar) {
cfang@1116	1183	md2 = (c - 1) - i;
cfang@1116	1184	}
cfang@1116	1185	}
cfang@1116	1186
cfang@1116	1187	result = string_indexOf(receiver, pat, o, cache, md2);
duke@435	1188	}
cfang@1116	1189
duke@435	1190	push(result);
duke@435	1191	return true;
duke@435	1192	}
duke@435	1193
duke@435	1194	//--------------------------pop_math_arg--------------------------------
duke@435	1195	// Pop a double argument to a math function from the stack
duke@435	1196	// rounding it if necessary.
duke@435	1197	Node * LibraryCallKit::pop_math_arg() {
duke@435	1198	Node *arg = pop_pair();
duke@435	1199	if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
duke@435	1200	arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
duke@435	1201	return arg;
duke@435	1202	}
duke@435	1203
duke@435	1204	//------------------------------inline_trig----------------------------------
duke@435	1205	// Inline sin/cos/tan instructions, if possible. If rounding is required, do
duke@435	1206	// argument reduction which will turn into a fast/slow diamond.
duke@435	1207	bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
duke@435	1208	_sp += arg_size(); // restore stack pointer
duke@435	1209	Node* arg = pop_math_arg();
duke@435	1210	Node* trig = NULL;
duke@435	1211
duke@435	1212	switch (id) {
duke@435	1213	case vmIntrinsics::_dsin:
duke@435	1214	trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
duke@435	1215	break;
duke@435	1216	case vmIntrinsics::_dcos:
duke@435	1217	trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
duke@435	1218	break;
duke@435	1219	case vmIntrinsics::_dtan:
duke@435	1220	trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
duke@435	1221	break;
duke@435	1222	default:
duke@435	1223	assert(false, "bad intrinsic was passed in");
duke@435	1224	return false;
duke@435	1225	}
duke@435	1226
duke@435	1227	// Rounding required? Check for argument reduction!
duke@435	1228	if( Matcher::strict_fp_requires_explicit_rounding ) {
duke@435	1229
duke@435	1230	static const double pi_4 = 0.7853981633974483;
duke@435	1231	static const double neg_pi_4 = -0.7853981633974483;
duke@435	1232	// pi/2 in 80-bit extended precision
duke@435	1233	// 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	1234	// -pi/2 in 80-bit extended precision
duke@435	1235	// 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	1236	// Cutoff value for using this argument reduction technique
duke@435	1237	//static const double pi_2_minus_epsilon = 1.564660403643354;
duke@435	1238	//static const double neg_pi_2_plus_epsilon = -1.564660403643354;
duke@435	1239
duke@435	1240	// Pseudocode for sin:
duke@435	1241	// if (x <= Math.PI / 4.0) {
duke@435	1242	// if (x >= -Math.PI / 4.0) return fsin(x);
duke@435	1243	// if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
duke@435	1244	// } else {
duke@435	1245	// if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
duke@435	1246	// }
duke@435	1247	// return StrictMath.sin(x);
duke@435	1248
duke@435	1249	// Pseudocode for cos:
duke@435	1250	// if (x <= Math.PI / 4.0) {
duke@435	1251	// if (x >= -Math.PI / 4.0) return fcos(x);
duke@435	1252	// if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
duke@435	1253	// } else {
duke@435	1254	// if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
duke@435	1255	// }
duke@435	1256	// return StrictMath.cos(x);
duke@435	1257
duke@435	1258	// Actually, sticking in an 80-bit Intel value into C2 will be tough; it
duke@435	1259	// requires a special machine instruction to load it. Instead we'll try
duke@435	1260	// the 'easy' case. If we really need the extra range +/- PI/2 we'll
duke@435	1261	// probably do the math inside the SIN encoding.
duke@435	1262
duke@435	1263	// Make the merge point
duke@435	1264	RegionNode *r = new (C, 3) RegionNode(3);
duke@435	1265	Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
duke@435	1266
duke@435	1267	// Flatten arg so we need only 1 test
duke@435	1268	Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
duke@435	1269	// Node for PI/4 constant
duke@435	1270	Node *pi4 = makecon(TypeD::make(pi_4));
duke@435	1271	// Check PI/4 : abs(arg)
duke@435	1272	Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
duke@435	1273	// Check: If PI/4 < abs(arg) then go slow
duke@435	1274	Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
duke@435	1275	// Branch either way
duke@435	1276	IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
duke@435	1277	set_control(opt_iff(r,iff));
duke@435	1278
duke@435	1279	// Set fast path result
duke@435	1280	phi->init_req(2,trig);
duke@435	1281
duke@435	1282	// Slow path - non-blocking leaf call
duke@435	1283	Node* call = NULL;
duke@435	1284	switch (id) {
duke@435	1285	case vmIntrinsics::_dsin:
duke@435	1286	call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
duke@435	1287	CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
duke@435	1288	"Sin", NULL, arg, top());
duke@435	1289	break;
duke@435	1290	case vmIntrinsics::_dcos:
duke@435	1291	call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
duke@435	1292	CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
duke@435	1293	"Cos", NULL, arg, top());
duke@435	1294	break;
duke@435	1295	case vmIntrinsics::_dtan:
duke@435	1296	call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
duke@435	1297	CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
duke@435	1298	"Tan", NULL, arg, top());
duke@435	1299	break;
duke@435	1300	}
duke@435	1301	assert(control()->in(0) == call, "");
duke@435	1302	Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
duke@435	1303	r->init_req(1,control());
duke@435	1304	phi->init_req(1,slow_result);
duke@435	1305
duke@435	1306	// Post-merge
duke@435	1307	set_control(_gvn.transform(r));
duke@435	1308	record_for_igvn(r);
duke@435	1309	trig = _gvn.transform(phi);
duke@435	1310
duke@435	1311	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	1312	}
duke@435	1313	// Push result back on JVM stack
duke@435	1314	push_pair(trig);
duke@435	1315	return true;
duke@435	1316	}
duke@435	1317
duke@435	1318	//------------------------------inline_sqrt-------------------------------------
duke@435	1319	// Inline square root instruction, if possible.
duke@435	1320	bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
duke@435	1321	assert(id == vmIntrinsics::_dsqrt, "Not square root");
duke@435	1322	_sp += arg_size(); // restore stack pointer
duke@435	1323	push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
duke@435	1324	return true;
duke@435	1325	}
duke@435	1326
duke@435	1327	//------------------------------inline_abs-------------------------------------
duke@435	1328	// Inline absolute value instruction, if possible.
duke@435	1329	bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
duke@435	1330	assert(id == vmIntrinsics::_dabs, "Not absolute value");
duke@435	1331	_sp += arg_size(); // restore stack pointer
duke@435	1332	push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
duke@435	1333	return true;
duke@435	1334	}
duke@435	1335
duke@435	1336	//------------------------------inline_exp-------------------------------------
duke@435	1337	// Inline exp instructions, if possible. The Intel hardware only misses
duke@435	1338	// really odd corner cases (+/- Infinity). Just uncommon-trap them.
duke@435	1339	bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
duke@435	1340	assert(id == vmIntrinsics::_dexp, "Not exp");
duke@435	1341
duke@435	1342	// If this inlining ever returned NaN in the past, we do not intrinsify it
duke@435	1343	// every again. NaN results requires StrictMath.exp handling.
duke@435	1344	if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@435	1345
duke@435	1346	// Do not intrinsify on older platforms which lack cmove.
duke@435	1347	if (ConditionalMoveLimit == 0) return false;
duke@435	1348
duke@435	1349	_sp += arg_size(); // restore stack pointer
duke@435	1350	Node *x = pop_math_arg();
duke@435	1351	Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
duke@435	1352
duke@435	1353	//-------------------
duke@435	1354	//result=(result.isNaN())? StrictMath::exp():result;
duke@435	1355	// Check: If isNaN() by checking result!=result? then go to Strict Math
duke@435	1356	Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
duke@435	1357	// Build the boolean node
duke@435	1358	Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
duke@435	1359
duke@435	1360	{ BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
duke@435	1361	// End the current control-flow path
duke@435	1362	push_pair(x);
duke@435	1363	// Math.exp intrinsic returned a NaN, which requires StrictMath.exp
duke@435	1364	// to handle. Recompile without intrinsifying Math.exp
duke@435	1365	uncommon_trap(Deoptimization::Reason_intrinsic,
duke@435	1366	Deoptimization::Action_make_not_entrant);
duke@435	1367	}
duke@435	1368
duke@435	1369	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	1370
duke@435	1371	push_pair(result);
duke@435	1372
duke@435	1373	return true;
duke@435	1374	}
duke@435	1375
duke@435	1376	//------------------------------inline_pow-------------------------------------
duke@435	1377	// Inline power instructions, if possible.
duke@435	1378	bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
duke@435	1379	assert(id == vmIntrinsics::_dpow, "Not pow");
duke@435	1380
duke@435	1381	// If this inlining ever returned NaN in the past, we do not intrinsify it
duke@435	1382	// every again. NaN results requires StrictMath.pow handling.
duke@435	1383	if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@435	1384
duke@435	1385	// Do not intrinsify on older platforms which lack cmove.
duke@435	1386	if (ConditionalMoveLimit == 0) return false;
duke@435	1387
duke@435	1388	// Pseudocode for pow
duke@435	1389	// if (x <= 0.0) {
duke@435	1390	// if ((double)((int)y)==y) { // if y is int
duke@435	1391	// result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
duke@435	1392	// } else {
duke@435	1393	// result = NaN;
duke@435	1394	// }
duke@435	1395	// } else {
duke@435	1396	// result = DPow(x,y);
duke@435	1397	// }
duke@435	1398	// if (result != result)? {
twisti@1040	1399	// uncommon_trap();
duke@435	1400	// }
duke@435	1401	// return result;
duke@435	1402
duke@435	1403	_sp += arg_size(); // restore stack pointer
duke@435	1404	Node* y = pop_math_arg();
duke@435	1405	Node* x = pop_math_arg();
duke@435	1406
duke@435	1407	Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
duke@435	1408
duke@435	1409	// Short form: if not top-level (i.e., Math.pow but inlining Math.pow
duke@435	1410	// inside of something) then skip the fancy tests and just check for
duke@435	1411	// NaN result.
duke@435	1412	Node *result = NULL;
duke@435	1413	if( jvms()->depth() >= 1 ) {
duke@435	1414	result = fast_result;
duke@435	1415	} else {
duke@435	1416
duke@435	1417	// Set the merge point for If node with condition of (x <= 0.0)
duke@435	1418	// There are four possible paths to region node and phi node
duke@435	1419	RegionNode *r = new (C, 4) RegionNode(4);
duke@435	1420	Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
duke@435	1421
duke@435	1422	// Build the first if node: if (x <= 0.0)
duke@435	1423	// Node for 0 constant
duke@435	1424	Node *zeronode = makecon(TypeD::ZERO);
duke@435	1425	// Check x:0
duke@435	1426	Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
duke@435	1427	// Check: If (x<=0) then go complex path
duke@435	1428	Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
duke@435	1429	// Branch either way
duke@435	1430	IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
duke@435	1431	Node *opt_test = _gvn.transform(if1);
duke@435	1432	//assert( opt_test->is_If(), "Expect an IfNode");
duke@435	1433	IfNode *opt_if1 = (IfNode*)opt_test;
duke@435	1434	// Fast path taken; set region slot 3
duke@435	1435	Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
duke@435	1436	r->init_req(3,fast_taken); // Capture fast-control
duke@435	1437
duke@435	1438	// Fast path not-taken, i.e. slow path
duke@435	1439	Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
duke@435	1440
duke@435	1441	// Set fast path result
duke@435	1442	Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
duke@435	1443	phi->init_req(3, fast_result);
duke@435	1444
duke@435	1445	// Complex path
duke@435	1446	// Build the second if node (if y is int)
duke@435	1447	// Node for (int)y
duke@435	1448	Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
duke@435	1449	// Node for (double)((int) y)
duke@435	1450	Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
duke@435	1451	// Check (double)((int) y) : y
duke@435	1452	Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
duke@435	1453	// Check if (y isn't int) then go to slow path
duke@435	1454
duke@435	1455	Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
twisti@1040	1456	// Branch either way
duke@435	1457	IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
duke@435	1458	Node *slow_path = opt_iff(r,if2); // Set region path 2
duke@435	1459
duke@435	1460	// Calculate DPow(abs(x), y)*(1 & (int)y)
duke@435	1461	// Node for constant 1
duke@435	1462	Node *conone = intcon(1);
duke@435	1463	// 1& (int)y
duke@435	1464	Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
duke@435	1465	// zero node
duke@435	1466	Node *conzero = intcon(0);
duke@435	1467	// Check (1&(int)y)==0?
duke@435	1468	Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
duke@435	1469	// Check if (1&(int)y)!=0?, if so the result is negative
duke@435	1470	Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
duke@435	1471	// abs(x)
duke@435	1472	Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
duke@435	1473	// abs(x)^y
duke@435	1474	Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
duke@435	1475	// -abs(x)^y
duke@435	1476	Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
duke@435	1477	// (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
duke@435	1478	Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
duke@435	1479	// Set complex path fast result
duke@435	1480	phi->init_req(2, signresult);
duke@435	1481
duke@435	1482	static const jlong nan_bits = CONST64(0x7ff8000000000000);
duke@435	1483	Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
duke@435	1484	r->init_req(1,slow_path);
duke@435	1485	phi->init_req(1,slow_result);
duke@435	1486
duke@435	1487	// Post merge
duke@435	1488	set_control(_gvn.transform(r));
duke@435	1489	record_for_igvn(r);
duke@435	1490	result=_gvn.transform(phi);
duke@435	1491	}
duke@435	1492
duke@435	1493	//-------------------
duke@435	1494	//result=(result.isNaN())? uncommon_trap():result;
duke@435	1495	// Check: If isNaN() by checking result!=result? then go to Strict Math
duke@435	1496	Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
duke@435	1497	// Build the boolean node
duke@435	1498	Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
duke@435	1499
duke@435	1500	{ BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
duke@435	1501	// End the current control-flow path
duke@435	1502	push_pair(x);
duke@435	1503	push_pair(y);
duke@435	1504	// Math.pow intrinsic returned a NaN, which requires StrictMath.pow
duke@435	1505	// to handle. Recompile without intrinsifying Math.pow.
duke@435	1506	uncommon_trap(Deoptimization::Reason_intrinsic,
duke@435	1507	Deoptimization::Action_make_not_entrant);
duke@435	1508	}
duke@435	1509
duke@435	1510	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	1511
duke@435	1512	push_pair(result);
duke@435	1513
duke@435	1514	return true;
duke@435	1515	}
duke@435	1516
duke@435	1517	//------------------------------inline_trans-------------------------------------
duke@435	1518	// Inline transcendental instructions, if possible. The Intel hardware gets
duke@435	1519	// these right, no funny corner cases missed.
duke@435	1520	bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
duke@435	1521	_sp += arg_size(); // restore stack pointer
duke@435	1522	Node* arg = pop_math_arg();
duke@435	1523	Node* trans = NULL;
duke@435	1524
duke@435	1525	switch (id) {
duke@435	1526	case vmIntrinsics::_dlog:
duke@435	1527	trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
duke@435	1528	break;
duke@435	1529	case vmIntrinsics::_dlog10:
duke@435	1530	trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
duke@435	1531	break;
duke@435	1532	default:
duke@435	1533	assert(false, "bad intrinsic was passed in");
duke@435	1534	return false;
duke@435	1535	}
duke@435	1536
duke@435	1537	// Push result back on JVM stack
duke@435	1538	push_pair(trans);
duke@435	1539	return true;
duke@435	1540	}
duke@435	1541
duke@435	1542	//------------------------------runtime_math-----------------------------
duke@435	1543	bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
duke@435	1544	Node* a = NULL;
duke@435	1545	Node* b = NULL;
duke@435	1546
duke@435	1547	assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
duke@435	1548	"must be (DD)D or (D)D type");
duke@435	1549
duke@435	1550	// Inputs
duke@435	1551	_sp += arg_size(); // restore stack pointer
duke@435	1552	if (call_type == OptoRuntime::Math_DD_D_Type()) {
duke@435	1553	b = pop_math_arg();
duke@435	1554	}
duke@435	1555	a = pop_math_arg();
duke@435	1556
duke@435	1557	const TypePtr* no_memory_effects = NULL;
duke@435	1558	Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
duke@435	1559	no_memory_effects,
duke@435	1560	a, top(), b, b ? top() : NULL);
duke@435	1561	Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
duke@435	1562	#ifdef ASSERT
duke@435	1563	Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
duke@435	1564	assert(value_top == top(), "second value must be top");
duke@435	1565	#endif
duke@435	1566
duke@435	1567	push_pair(value);
duke@435	1568	return true;
duke@435	1569	}
duke@435	1570
duke@435	1571	//------------------------------inline_math_native-----------------------------
duke@435	1572	bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
duke@435	1573	switch (id) {
duke@435	1574	// These intrinsics are not properly supported on all hardware
duke@435	1575	case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
duke@435	1576	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
duke@435	1577	case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
duke@435	1578	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
duke@435	1579	case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
duke@435	1580	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
duke@435	1581
duke@435	1582	case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
duke@435	1583	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
duke@435	1584	case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
duke@435	1585	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
duke@435	1586
duke@435	1587	// These intrinsics are supported on all hardware
duke@435	1588	case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
duke@435	1589	case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
duke@435	1590
duke@435	1591	// These intrinsics don't work on X86. The ad implementation doesn't
duke@435	1592	// handle NaN's properly. Instead of returning infinity, the ad
duke@435	1593	// implementation returns a NaN on overflow. See bug: 6304089
duke@435	1594	// Once the ad implementations are fixed, change the code below
duke@435	1595	// to match the intrinsics above
duke@435	1596
duke@435	1597	case vmIntrinsics::_dexp: return
duke@435	1598	runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
duke@435	1599	case vmIntrinsics::_dpow: return
duke@435	1600	runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
duke@435	1601
duke@435	1602	// These intrinsics are not yet correctly implemented
duke@435	1603	case vmIntrinsics::_datan2:
duke@435	1604	return false;
duke@435	1605
duke@435	1606	default:
duke@435	1607	ShouldNotReachHere();
duke@435	1608	return false;
duke@435	1609	}
duke@435	1610	}
duke@435	1611
duke@435	1612	static bool is_simple_name(Node* n) {
duke@435	1613	return (n->req() == 1 // constant
duke@435	1614	|| (n->is_Type() && n->as_Type()->type()->singleton())
duke@435	1615	|| n->is_Proj() // parameter or return value
duke@435	1616	|| n->is_Phi() // local of some sort
duke@435	1617	);
duke@435	1618	}
duke@435	1619
duke@435	1620	//----------------------------inline_min_max-----------------------------------
duke@435	1621	bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
duke@435	1622	push(generate_min_max(id, argument(0), argument(1)));
duke@435	1623
duke@435	1624	return true;
duke@435	1625	}
duke@435	1626
duke@435	1627	Node*
duke@435	1628	LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
duke@435	1629	// These are the candidate return value:
duke@435	1630	Node* xvalue = x0;
duke@435	1631	Node* yvalue = y0;
duke@435	1632
duke@435	1633	if (xvalue == yvalue) {
duke@435	1634	return xvalue;
duke@435	1635	}
duke@435	1636
duke@435	1637	bool want_max = (id == vmIntrinsics::_max);
duke@435	1638
duke@435	1639	const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
duke@435	1640	const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
duke@435	1641	if (txvalue == NULL || tyvalue == NULL) return top();
duke@435	1642	// This is not really necessary, but it is consistent with a
duke@435	1643	// hypothetical MaxINode::Value method:
duke@435	1644	int widen = MAX2(txvalue->_widen, tyvalue->_widen);
duke@435	1645
duke@435	1646	// %%% This folding logic should (ideally) be in a different place.
duke@435	1647	// Some should be inside IfNode, and there to be a more reliable
duke@435	1648	// transformation of ?: style patterns into cmoves. We also want
duke@435	1649	// more powerful optimizations around cmove and min/max.
duke@435	1650
duke@435	1651	// Try to find a dominating comparison of these guys.
duke@435	1652	// It can simplify the index computation for Arrays.copyOf
duke@435	1653	// and similar uses of System.arraycopy.
duke@435	1654	// First, compute the normalized version of CmpI(x, y).
duke@435	1655	int cmp_op = Op_CmpI;
duke@435	1656	Node* xkey = xvalue;
duke@435	1657	Node* ykey = yvalue;
duke@435	1658	Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
duke@435	1659	if (ideal_cmpxy->is_Cmp()) {
duke@435	1660	// E.g., if we have CmpI(length - offset, count),
duke@435	1661	// it might idealize to CmpI(length, count + offset)
duke@435	1662	cmp_op = ideal_cmpxy->Opcode();
duke@435	1663	xkey = ideal_cmpxy->in(1);
duke@435	1664	ykey = ideal_cmpxy->in(2);
duke@435	1665	}
duke@435	1666
duke@435	1667	// Start by locating any relevant comparisons.
duke@435	1668	Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
duke@435	1669	Node* cmpxy = NULL;
duke@435	1670	Node* cmpyx = NULL;
duke@435	1671	for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
duke@435	1672	Node* cmp = start_from->fast_out(k);
duke@435	1673	if (cmp->outcnt() > 0 && // must have prior uses
duke@435	1674	cmp->in(0) == NULL && // must be context-independent
duke@435	1675	cmp->Opcode() == cmp_op) { // right kind of compare
duke@435	1676	if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
duke@435	1677	if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
duke@435	1678	}
duke@435	1679	}
duke@435	1680
duke@435	1681	const int NCMPS = 2;
duke@435	1682	Node* cmps[NCMPS] = { cmpxy, cmpyx };
duke@435	1683	int cmpn;
duke@435	1684	for (cmpn = 0; cmpn < NCMPS; cmpn++) {
duke@435	1685	if (cmps[cmpn] != NULL) break; // find a result
duke@435	1686	}
duke@435	1687	if (cmpn < NCMPS) {
duke@435	1688	// Look for a dominating test that tells us the min and max.
duke@435	1689	int depth = 0; // Limit search depth for speed
duke@435	1690	Node* dom = control();
duke@435	1691	for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
duke@435	1692	if (++depth >= 100) break;
duke@435	1693	Node* ifproj = dom;
duke@435	1694	if (!ifproj->is_Proj()) continue;
duke@435	1695	Node* iff = ifproj->in(0);
duke@435	1696	if (!iff->is_If()) continue;
duke@435	1697	Node* bol = iff->in(1);
duke@435	1698	if (!bol->is_Bool()) continue;
duke@435	1699	Node* cmp = bol->in(1);
duke@435	1700	if (cmp == NULL) continue;
duke@435	1701	for (cmpn = 0; cmpn < NCMPS; cmpn++)
duke@435	1702	if (cmps[cmpn] == cmp) break;
duke@435	1703	if (cmpn == NCMPS) continue;
duke@435	1704	BoolTest::mask btest = bol->as_Bool()->_test._test;
duke@435	1705	if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
duke@435	1706	if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
duke@435	1707	// At this point, we know that 'x btest y' is true.
duke@435	1708	switch (btest) {
duke@435	1709	case BoolTest::eq:
duke@435	1710	// They are proven equal, so we can collapse the min/max.
duke@435	1711	// Either value is the answer. Choose the simpler.
duke@435	1712	if (is_simple_name(yvalue) && !is_simple_name(xvalue))
duke@435	1713	return yvalue;
duke@435	1714	return xvalue;
duke@435	1715	case BoolTest::lt: // x < y
duke@435	1716	case BoolTest::le: // x <= y
duke@435	1717	return (want_max ? yvalue : xvalue);
duke@435	1718	case BoolTest::gt: // x > y
duke@435	1719	case BoolTest::ge: // x >= y
duke@435	1720	return (want_max ? xvalue : yvalue);
duke@435	1721	}
duke@435	1722	}
duke@435	1723	}
duke@435	1724
duke@435	1725	// We failed to find a dominating test.
duke@435	1726	// Let's pick a test that might GVN with prior tests.
duke@435	1727	Node* best_bol = NULL;
duke@435	1728	BoolTest::mask best_btest = BoolTest::illegal;
duke@435	1729	for (cmpn = 0; cmpn < NCMPS; cmpn++) {
duke@435	1730	Node* cmp = cmps[cmpn];
duke@435	1731	if (cmp == NULL) continue;
duke@435	1732	for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
duke@435	1733	Node* bol = cmp->fast_out(j);
duke@435	1734	if (!bol->is_Bool()) continue;
duke@435	1735	BoolTest::mask btest = bol->as_Bool()->_test._test;
duke@435	1736	if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
duke@435	1737	if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
duke@435	1738	if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
duke@435	1739	best_bol = bol->as_Bool();
duke@435	1740	best_btest = btest;
duke@435	1741	}
duke@435	1742	}
duke@435	1743	}
duke@435	1744
duke@435	1745	Node* answer_if_true = NULL;
duke@435	1746	Node* answer_if_false = NULL;
duke@435	1747	switch (best_btest) {
duke@435	1748	default:
duke@435	1749	if (cmpxy == NULL)
duke@435	1750	cmpxy = ideal_cmpxy;
duke@435	1751	best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
duke@435	1752	// and fall through:
duke@435	1753	case BoolTest::lt: // x < y
duke@435	1754	case BoolTest::le: // x <= y
duke@435	1755	answer_if_true = (want_max ? yvalue : xvalue);
duke@435	1756	answer_if_false = (want_max ? xvalue : yvalue);
duke@435	1757	break;
duke@435	1758	case BoolTest::gt: // x > y
duke@435	1759	case BoolTest::ge: // x >= y
duke@435	1760	answer_if_true = (want_max ? xvalue : yvalue);
duke@435	1761	answer_if_false = (want_max ? yvalue : xvalue);
duke@435	1762	break;
duke@435	1763	}
duke@435	1764
duke@435	1765	jint hi, lo;
duke@435	1766	if (want_max) {
duke@435	1767	// We can sharpen the minimum.
duke@435	1768	hi = MAX2(txvalue->_hi, tyvalue->_hi);
duke@435	1769	lo = MAX2(txvalue->_lo, tyvalue->_lo);
duke@435	1770	} else {
duke@435	1771	// We can sharpen the maximum.
duke@435	1772	hi = MIN2(txvalue->_hi, tyvalue->_hi);
duke@435	1773	lo = MIN2(txvalue->_lo, tyvalue->_lo);
duke@435	1774	}
duke@435	1775
duke@435	1776	// Use a flow-free graph structure, to avoid creating excess control edges
duke@435	1777	// which could hinder other optimizations.
duke@435	1778	// Since Math.min/max is often used with arraycopy, we want
duke@435	1779	// tightly_coupled_allocation to be able to see beyond min/max expressions.
duke@435	1780	Node* cmov = CMoveNode::make(C, NULL, best_bol,
duke@435	1781	answer_if_false, answer_if_true,
duke@435	1782	TypeInt::make(lo, hi, widen));
duke@435	1783
duke@435	1784	return _gvn.transform(cmov);
duke@435	1785
duke@435	1786	/*
duke@435	1787	// This is not as desirable as it may seem, since Min and Max
duke@435	1788	// nodes do not have a full set of optimizations.
duke@435	1789	// And they would interfere, anyway, with 'if' optimizations
duke@435	1790	// and with CMoveI canonical forms.
duke@435	1791	switch (id) {
duke@435	1792	case vmIntrinsics::_min:
duke@435	1793	result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
duke@435	1794	case vmIntrinsics::_max:
duke@435	1795	result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
duke@435	1796	default:
duke@435	1797	ShouldNotReachHere();
duke@435	1798	}
duke@435	1799	*/
duke@435	1800	}
duke@435	1801
duke@435	1802	inline int
duke@435	1803	LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
duke@435	1804	const TypePtr* base_type = TypePtr::NULL_PTR;
duke@435	1805	if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
duke@435	1806	if (base_type == NULL) {
duke@435	1807	// Unknown type.
duke@435	1808	return Type::AnyPtr;
duke@435	1809	} else if (base_type == TypePtr::NULL_PTR) {
duke@435	1810	// Since this is a NULL+long form, we have to switch to a rawptr.
duke@435	1811	base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
duke@435	1812	offset = MakeConX(0);
duke@435	1813	return Type::RawPtr;
duke@435	1814	} else if (base_type->base() == Type::RawPtr) {
duke@435	1815	return Type::RawPtr;
duke@435	1816	} else if (base_type->isa_oopptr()) {
duke@435	1817	// Base is never null => always a heap address.
duke@435	1818	if (base_type->ptr() == TypePtr::NotNull) {
duke@435	1819	return Type::OopPtr;
duke@435	1820	}
duke@435	1821	// Offset is small => always a heap address.
duke@435	1822	const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
duke@435	1823	if (offset_type != NULL &&
duke@435	1824	base_type->offset() == 0 && // (should always be?)
duke@435	1825	offset_type->_lo >= 0 &&
duke@435	1826	!MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
duke@435	1827	return Type::OopPtr;
duke@435	1828	}
duke@435	1829	// Otherwise, it might either be oop+off or NULL+addr.
duke@435	1830	return Type::AnyPtr;
duke@435	1831	} else {
duke@435	1832	// No information:
duke@435	1833	return Type::AnyPtr;
duke@435	1834	}
duke@435	1835	}
duke@435	1836
duke@435	1837	inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
duke@435	1838	int kind = classify_unsafe_addr(base, offset);
duke@435	1839	if (kind == Type::RawPtr) {
duke@435	1840	return basic_plus_adr(top(), base, offset);
duke@435	1841	} else {
duke@435	1842	return basic_plus_adr(base, offset);
duke@435	1843	}
duke@435	1844	}
duke@435	1845
twisti@1210	1846	//-------------------inline_numberOfLeadingZeros_int/long-----------------------
twisti@1210	1847	// inline int Integer.numberOfLeadingZeros(int)
twisti@1210	1848	// inline int Long.numberOfLeadingZeros(long)
twisti@1210	1849	bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
twisti@1210	1850	assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
twisti@1210	1851	if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
twisti@1210	1852	if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
twisti@1210	1853	_sp += arg_size(); // restore stack pointer
twisti@1210	1854	switch (id) {
twisti@1210	1855	case vmIntrinsics::_numberOfLeadingZeros_i:
twisti@1210	1856	push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
twisti@1210	1857	break;
twisti@1210	1858	case vmIntrinsics::_numberOfLeadingZeros_l:
twisti@1210	1859	push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
twisti@1210	1860	break;
twisti@1210	1861	default:
twisti@1210	1862	ShouldNotReachHere();
twisti@1210	1863	}
twisti@1210	1864	return true;
twisti@1210	1865	}
twisti@1210	1866
twisti@1210	1867	//-------------------inline_numberOfTrailingZeros_int/long----------------------
twisti@1210	1868	// inline int Integer.numberOfTrailingZeros(int)
twisti@1210	1869	// inline int Long.numberOfTrailingZeros(long)
twisti@1210	1870	bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
twisti@1210	1871	assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
twisti@1210	1872	if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
twisti@1210	1873	if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
twisti@1210	1874	_sp += arg_size(); // restore stack pointer
twisti@1210	1875	switch (id) {
twisti@1210	1876	case vmIntrinsics::_numberOfTrailingZeros_i:
twisti@1210	1877	push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
twisti@1210	1878	break;
twisti@1210	1879	case vmIntrinsics::_numberOfTrailingZeros_l:
twisti@1210	1880	push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
twisti@1210	1881	break;
twisti@1210	1882	default:
twisti@1210	1883	ShouldNotReachHere();
twisti@1210	1884	}
twisti@1210	1885	return true;
twisti@1210	1886	}
twisti@1210	1887
twisti@1078	1888	//----------------------------inline_bitCount_int/long-----------------------
twisti@1078	1889	// inline int Integer.bitCount(int)
twisti@1078	1890	// inline int Long.bitCount(long)
twisti@1078	1891	bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
twisti@1078	1892	assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
twisti@1078	1893	if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
twisti@1078	1894	if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
twisti@1078	1895	_sp += arg_size(); // restore stack pointer
twisti@1078	1896	switch (id) {
twisti@1078	1897	case vmIntrinsics::_bitCount_i:
twisti@1078	1898	push(_gvn.transform(new (C, 2) PopCountINode(pop())));
twisti@1078	1899	break;
twisti@1078	1900	case vmIntrinsics::_bitCount_l:
twisti@1078	1901	push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
twisti@1078	1902	break;
twisti@1078	1903	default:
twisti@1078	1904	ShouldNotReachHere();
twisti@1078	1905	}
twisti@1078	1906	return true;
twisti@1078	1907	}
twisti@1078	1908
duke@435	1909	//----------------------------inline_reverseBytes_int/long-------------------
twisti@1040	1910	// inline Integer.reverseBytes(int)
twisti@1040	1911	// inline Long.reverseBytes(long)
duke@435	1912	bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
duke@435	1913	assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes");
duke@435	1914	if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
duke@435	1915	if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
duke@435	1916	_sp += arg_size(); // restore stack pointer
duke@435	1917	switch (id) {
duke@435	1918	case vmIntrinsics::_reverseBytes_i:
duke@435	1919	push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
duke@435	1920	break;
duke@435	1921	case vmIntrinsics::_reverseBytes_l:
duke@435	1922	push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
duke@435	1923	break;
duke@435	1924	default:
duke@435	1925	;
duke@435	1926	}
duke@435	1927	return true;
duke@435	1928	}
duke@435	1929
duke@435	1930	//----------------------------inline_unsafe_access----------------------------
duke@435	1931
duke@435	1932	const static BasicType T_ADDRESS_HOLDER = T_LONG;
duke@435	1933
duke@435	1934	// Interpret Unsafe.fieldOffset cookies correctly:
duke@435	1935	extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
duke@435	1936
duke@435	1937	bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
duke@435	1938	if (callee()->is_static()) return false; // caller must have the capability!
duke@435	1939
duke@435	1940	#ifndef PRODUCT
duke@435	1941	{
duke@435	1942	ResourceMark rm;
duke@435	1943	// Check the signatures.
duke@435	1944	ciSignature* sig = signature();
duke@435	1945	#ifdef ASSERT
duke@435	1946	if (!is_store) {
duke@435	1947	// Object getObject(Object base, int/long offset), etc.
duke@435	1948	BasicType rtype = sig->return_type()->basic_type();
duke@435	1949	if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
duke@435	1950	rtype = T_ADDRESS; // it is really a C void*
duke@435	1951	assert(rtype == type, "getter must return the expected value");
duke@435	1952	if (!is_native_ptr) {
duke@435	1953	assert(sig->count() == 2, "oop getter has 2 arguments");
duke@435	1954	assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
duke@435	1955	assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
duke@435	1956	} else {
duke@435	1957	assert(sig->count() == 1, "native getter has 1 argument");
duke@435	1958	assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
duke@435	1959	}
duke@435	1960	} else {
duke@435	1961	// void putObject(Object base, int/long offset, Object x), etc.
duke@435	1962	assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
duke@435	1963	if (!is_native_ptr) {
duke@435	1964	assert(sig->count() == 3, "oop putter has 3 arguments");
duke@435	1965	assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
duke@435	1966	assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
duke@435	1967	} else {
duke@435	1968	assert(sig->count() == 2, "native putter has 2 arguments");
duke@435	1969	assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
duke@435	1970	}
duke@435	1971	BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
duke@435	1972	if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
duke@435	1973	vtype = T_ADDRESS; // it is really a C void*
duke@435	1974	assert(vtype == type, "putter must accept the expected value");
duke@435	1975	}
duke@435	1976	#endif // ASSERT
duke@435	1977	}
duke@435	1978	#endif //PRODUCT
duke@435	1979
duke@435	1980	C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@435	1981
duke@435	1982	int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
duke@435	1983
duke@435	1984	// Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
duke@435	1985	int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
duke@435	1986
duke@435	1987	debug_only(int saved_sp = _sp);
duke@435	1988	_sp += nargs;
duke@435	1989
duke@435	1990	Node* val;
duke@435	1991	debug_only(val = (Node*)(uintptr_t)-1);
duke@435	1992
duke@435	1993
duke@435	1994	if (is_store) {
duke@435	1995	// Get the value being stored. (Pop it first; it was pushed last.)
duke@435	1996	switch (type) {
duke@435	1997	case T_DOUBLE:
duke@435	1998	case T_LONG:
duke@435	1999	case T_ADDRESS:
duke@435	2000	val = pop_pair();
duke@435	2001	break;
duke@435	2002	default:
duke@435	2003	val = pop();
duke@435	2004	}
duke@435	2005	}
duke@435	2006
duke@435	2007	// Build address expression. See the code in inline_unsafe_prefetch.
duke@435	2008	Node *adr;
duke@435	2009	Node *heap_base_oop = top();
duke@435	2010	if (!is_native_ptr) {
duke@435	2011	// The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
duke@435	2012	Node* offset = pop_pair();
duke@435	2013	// The base is either a Java object or a value produced by Unsafe.staticFieldBase
duke@435	2014	Node* base = pop();
duke@435	2015	// We currently rely on the cookies produced by Unsafe.xxxFieldOffset
duke@435	2016	// to be plain byte offsets, which are also the same as those accepted
duke@435	2017	// by oopDesc::field_base.
duke@435	2018	assert(Unsafe_field_offset_to_byte_offset(11) == 11,
duke@435	2019	"fieldOffset must be byte-scaled");
duke@435	2020	// 32-bit machines ignore the high half!
duke@435	2021	offset = ConvL2X(offset);
duke@435	2022	adr = make_unsafe_address(base, offset);
duke@435	2023	heap_base_oop = base;
duke@435	2024	} else {
duke@435	2025	Node* ptr = pop_pair();
duke@435	2026	// Adjust Java long to machine word:
duke@435	2027	ptr = ConvL2X(ptr);
duke@435	2028	adr = make_unsafe_address(NULL, ptr);
duke@435	2029	}
duke@435	2030
duke@435	2031	// Pop receiver last: it was pushed first.
duke@435	2032	Node *receiver = pop();
duke@435	2033
duke@435	2034	assert(saved_sp == _sp, "must have correct argument count");
duke@435	2035
duke@435	2036	const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
duke@435	2037
duke@435	2038	// First guess at the value type.
duke@435	2039	const Type *value_type = Type::get_const_basic_type(type);
duke@435	2040
duke@435	2041	// Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
duke@435	2042	// there was not enough information to nail it down.
duke@435	2043	Compile::AliasType* alias_type = C->alias_type(adr_type);
duke@435	2044	assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
duke@435	2045
duke@435	2046	// We will need memory barriers unless we can determine a unique
duke@435	2047	// alias category for this reference. (Note: If for some reason
duke@435	2048	// the barriers get omitted and the unsafe reference begins to "pollute"
duke@435	2049	// the alias analysis of the rest of the graph, either Compile::can_alias
duke@435	2050	// or Compile::must_alias will throw a diagnostic assert.)
duke@435	2051	bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
duke@435	2052
duke@435	2053	if (!is_store && type == T_OBJECT) {
duke@435	2054	// Attempt to infer a sharper value type from the offset and base type.
duke@435	2055	ciKlass* sharpened_klass = NULL;
duke@435	2056
duke@435	2057	// See if it is an instance field, with an object type.
duke@435	2058	if (alias_type->field() != NULL) {
duke@435	2059	assert(!is_native_ptr, "native pointer op cannot use a java address");
duke@435	2060	if (alias_type->field()->type()->is_klass()) {
duke@435	2061	sharpened_klass = alias_type->field()->type()->as_klass();
duke@435	2062	}
duke@435	2063	}
duke@435	2064
duke@435	2065	// See if it is a narrow oop array.
duke@435	2066	if (adr_type->isa_aryptr()) {
coleenp@548	2067	if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes(type)) {
duke@435	2068	const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
duke@435	2069	if (elem_type != NULL) {
duke@435	2070	sharpened_klass = elem_type->klass();
duke@435	2071	}
duke@435	2072	}
duke@435	2073	}
duke@435	2074
duke@435	2075	if (sharpened_klass != NULL) {
duke@435	2076	const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
duke@435	2077
duke@435	2078	// Sharpen the value type.
duke@435	2079	value_type = tjp;
duke@435	2080
duke@435	2081	#ifndef PRODUCT
duke@435	2082	if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
duke@435	2083	tty->print(" from base type: "); adr_type->dump();
duke@435	2084	tty->print(" sharpened value: "); value_type->dump();
duke@435	2085	}
duke@435	2086	#endif
duke@435	2087	}
duke@435	2088	}
duke@435	2089
duke@435	2090	// Null check on self without removing any arguments. The argument
duke@435	2091	// null check technically happens in the wrong place, which can lead to
duke@435	2092	// invalid stack traces when the primitive is inlined into a method
duke@435	2093	// which handles NullPointerExceptions.
duke@435	2094	_sp += nargs;
duke@435	2095	do_null_check(receiver, T_OBJECT);
duke@435	2096	_sp -= nargs;
duke@435	2097	if (stopped()) {
duke@435	2098	return true;
duke@435	2099	}
duke@435	2100	// Heap pointers get a null-check from the interpreter,
duke@435	2101	// as a courtesy. However, this is not guaranteed by Unsafe,
duke@435	2102	// and it is not possible to fully distinguish unintended nulls
duke@435	2103	// from intended ones in this API.
duke@435	2104
duke@435	2105	if (is_volatile) {
duke@435	2106	// We need to emit leading and trailing CPU membars (see below) in
duke@435	2107	// addition to memory membars when is_volatile. This is a little
duke@435	2108	// too strong, but avoids the need to insert per-alias-type
duke@435	2109	// volatile membars (for stores; compare Parse::do_put_xxx), which
twisti@1040	2110	// we cannot do effectively here because we probably only have a
duke@435	2111	// rough approximation of type.
duke@435	2112	need_mem_bar = true;
duke@435	2113	// For Stores, place a memory ordering barrier now.
duke@435	2114	if (is_store)
duke@435	2115	insert_mem_bar(Op_MemBarRelease);
duke@435	2116	}
duke@435	2117
duke@435	2118	// Memory barrier to prevent normal and 'unsafe' accesses from
duke@435	2119	// bypassing each other. Happens after null checks, so the
duke@435	2120	// exception paths do not take memory state from the memory barrier,
duke@435	2121	// so there's no problems making a strong assert about mixing users
duke@435	2122	// of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
duke@435	2123	// around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
duke@435	2124	if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2125
duke@435	2126	if (!is_store) {
duke@435	2127	Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
duke@435	2128	// load value and push onto stack
duke@435	2129	switch (type) {
duke@435	2130	case T_BOOLEAN:
duke@435	2131	case T_CHAR:
duke@435	2132	case T_BYTE:
duke@435	2133	case T_SHORT:
duke@435	2134	case T_INT:
duke@435	2135	case T_FLOAT:
duke@435	2136	case T_OBJECT:
duke@435	2137	push( p );
duke@435	2138	break;
duke@435	2139	case T_ADDRESS:
duke@435	2140	// Cast to an int type.
duke@435	2141	p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
duke@435	2142	p = ConvX2L(p);
duke@435	2143	push_pair(p);
duke@435	2144	break;
duke@435	2145	case T_DOUBLE:
duke@435	2146	case T_LONG:
duke@435	2147	push_pair( p );
duke@435	2148	break;
duke@435	2149	default: ShouldNotReachHere();
duke@435	2150	}
duke@435	2151	} else {
duke@435	2152	// place effect of store into memory
duke@435	2153	switch (type) {
duke@435	2154	case T_DOUBLE:
duke@435	2155	val = dstore_rounding(val);
duke@435	2156	break;
duke@435	2157	case T_ADDRESS:
duke@435	2158	// Repackage the long as a pointer.
duke@435	2159	val = ConvL2X(val);
duke@435	2160	val = _gvn.transform( new (C, 2) CastX2PNode(val) );
duke@435	2161	break;
duke@435	2162	}
duke@435	2163
duke@435	2164	if (type != T_OBJECT ) {
duke@435	2165	(void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
duke@435	2166	} else {
duke@435	2167	// Possibly an oop being stored to Java heap or native memory
duke@435	2168	if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
duke@435	2169	// oop to Java heap.
never@1260	2170	(void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
duke@435	2171	} else {
duke@435	2172	// We can't tell at compile time if we are storing in the Java heap or outside
duke@435	2173	// of it. So we need to emit code to conditionally do the proper type of
duke@435	2174	// store.
duke@435	2175
kvn@1286	2176	IdealKit ideal(gvn(), control(), merged_memory());
kvn@1286	2177	#define __ ideal.
duke@435	2178	// QQQ who knows what probability is here??
kvn@1286	2179	__ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
kvn@1286	2180	// Sync IdealKit and graphKit.
kvn@1286	2181	set_all_memory( __ merged_memory());
kvn@1286	2182	set_control(__ ctrl());
kvn@1286	2183	Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
kvn@1286	2184	// Update IdealKit memory.
kvn@1286	2185	__ set_all_memory(merged_memory());
kvn@1286	2186	__ set_ctrl(control());
kvn@1286	2187	} __ else_(); {
kvn@1286	2188	__ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
kvn@1286	2189	} __ end_if();
kvn@1286	2190	// Final sync IdealKit and GraphKit.
kvn@1286	2191	sync_kit(ideal);
kvn@1286	2192	#undef __
duke@435	2193	}
duke@435	2194	}
duke@435	2195	}
duke@435	2196
duke@435	2197	if (is_volatile) {
duke@435	2198	if (!is_store)
duke@435	2199	insert_mem_bar(Op_MemBarAcquire);
duke@435	2200	else
duke@435	2201	insert_mem_bar(Op_MemBarVolatile);
duke@435	2202	}
duke@435	2203
duke@435	2204	if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2205
duke@435	2206	return true;
duke@435	2207	}
duke@435	2208
duke@435	2209	//----------------------------inline_unsafe_prefetch----------------------------
duke@435	2210
duke@435	2211	bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
duke@435	2212	#ifndef PRODUCT
duke@435	2213	{
duke@435	2214	ResourceMark rm;
duke@435	2215	// Check the signatures.
duke@435	2216	ciSignature* sig = signature();
duke@435	2217	#ifdef ASSERT
duke@435	2218	// Object getObject(Object base, int/long offset), etc.
duke@435	2219	BasicType rtype = sig->return_type()->basic_type();
duke@435	2220	if (!is_native_ptr) {
duke@435	2221	assert(sig->count() == 2, "oop prefetch has 2 arguments");
duke@435	2222	assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
duke@435	2223	assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
duke@435	2224	} else {
duke@435	2225	assert(sig->count() == 1, "native prefetch has 1 argument");
duke@435	2226	assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
duke@435	2227	}
duke@435	2228	#endif // ASSERT
duke@435	2229	}
duke@435	2230	#endif // !PRODUCT
duke@435	2231
duke@435	2232	C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@435	2233
duke@435	2234	// Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
duke@435	2235	int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
duke@435	2236
duke@435	2237	debug_only(int saved_sp = _sp);
duke@435	2238	_sp += nargs;
duke@435	2239
duke@435	2240	// Build address expression. See the code in inline_unsafe_access.
duke@435	2241	Node *adr;
duke@435	2242	if (!is_native_ptr) {
duke@435	2243	// The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
duke@435	2244	Node* offset = pop_pair();
duke@435	2245	// The base is either a Java object or a value produced by Unsafe.staticFieldBase
duke@435	2246	Node* base = pop();
duke@435	2247	// We currently rely on the cookies produced by Unsafe.xxxFieldOffset
duke@435	2248	// to be plain byte offsets, which are also the same as those accepted
duke@435	2249	// by oopDesc::field_base.
duke@435	2250	assert(Unsafe_field_offset_to_byte_offset(11) == 11,
duke@435	2251	"fieldOffset must be byte-scaled");
duke@435	2252	// 32-bit machines ignore the high half!
duke@435	2253	offset = ConvL2X(offset);
duke@435	2254	adr = make_unsafe_address(base, offset);
duke@435	2255	} else {
duke@435	2256	Node* ptr = pop_pair();
duke@435	2257	// Adjust Java long to machine word:
duke@435	2258	ptr = ConvL2X(ptr);
duke@435	2259	adr = make_unsafe_address(NULL, ptr);
duke@435	2260	}
duke@435	2261
duke@435	2262	if (is_static) {
duke@435	2263	assert(saved_sp == _sp, "must have correct argument count");
duke@435	2264	} else {
duke@435	2265	// Pop receiver last: it was pushed first.
duke@435	2266	Node *receiver = pop();
duke@435	2267	assert(saved_sp == _sp, "must have correct argument count");
duke@435	2268
duke@435	2269	// Null check on self without removing any arguments. The argument
duke@435	2270	// null check technically happens in the wrong place, which can lead to
duke@435	2271	// invalid stack traces when the primitive is inlined into a method
duke@435	2272	// which handles NullPointerExceptions.
duke@435	2273	_sp += nargs;
duke@435	2274	do_null_check(receiver, T_OBJECT);
duke@435	2275	_sp -= nargs;
duke@435	2276	if (stopped()) {
duke@435	2277	return true;
duke@435	2278	}
duke@435	2279	}
duke@435	2280
duke@435	2281	// Generate the read or write prefetch
duke@435	2282	Node *prefetch;
duke@435	2283	if (is_store) {
duke@435	2284	prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
duke@435	2285	} else {
duke@435	2286	prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
duke@435	2287	}
duke@435	2288	prefetch->init_req(0, control());
duke@435	2289	set_i_o(_gvn.transform(prefetch));
duke@435	2290
duke@435	2291	return true;
duke@435	2292	}
duke@435	2293
duke@435	2294	//----------------------------inline_unsafe_CAS----------------------------
duke@435	2295
duke@435	2296	bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
duke@435	2297	// This basic scheme here is the same as inline_unsafe_access, but
duke@435	2298	// differs in enough details that combining them would make the code
duke@435	2299	// overly confusing. (This is a true fact! I originally combined
duke@435	2300	// them, but even I was confused by it!) As much code/comments as
duke@435	2301	// possible are retained from inline_unsafe_access though to make
twisti@1040	2302	// the correspondences clearer. - dl
duke@435	2303
duke@435	2304	if (callee()->is_static()) return false; // caller must have the capability!
duke@435	2305
duke@435	2306	#ifndef PRODUCT
duke@435	2307	{
duke@435	2308	ResourceMark rm;
duke@435	2309	// Check the signatures.
duke@435	2310	ciSignature* sig = signature();
duke@435	2311	#ifdef ASSERT
duke@435	2312	BasicType rtype = sig->return_type()->basic_type();
duke@435	2313	assert(rtype == T_BOOLEAN, "CAS must return boolean");
duke@435	2314	assert(sig->count() == 4, "CAS has 4 arguments");
duke@435	2315	assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
duke@435	2316	assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
duke@435	2317	#endif // ASSERT
duke@435	2318	}
duke@435	2319	#endif //PRODUCT
duke@435	2320
duke@435	2321	// number of stack slots per value argument (1 or 2)
duke@435	2322	int type_words = type2size[type];
duke@435	2323
duke@435	2324	// Cannot inline wide CAS on machines that don't support it natively
kvn@464	2325	if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
duke@435	2326	return false;
duke@435	2327
duke@435	2328	C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@435	2329
duke@435	2330	// Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
duke@435	2331	int nargs = 1 + 1 + 2 + type_words + type_words;
duke@435	2332
duke@435	2333	// pop arguments: newval, oldval, offset, base, and receiver
duke@435	2334	debug_only(int saved_sp = _sp);
duke@435	2335	_sp += nargs;
duke@435	2336	Node* newval = (type_words == 1) ? pop() : pop_pair();
duke@435	2337	Node* oldval = (type_words == 1) ? pop() : pop_pair();
duke@435	2338	Node *offset = pop_pair();
duke@435	2339	Node *base = pop();
duke@435	2340	Node *receiver = pop();
duke@435	2341	assert(saved_sp == _sp, "must have correct argument count");
duke@435	2342
duke@435	2343	// Null check receiver.
duke@435	2344	_sp += nargs;
duke@435	2345	do_null_check(receiver, T_OBJECT);
duke@435	2346	_sp -= nargs;
duke@435	2347	if (stopped()) {
duke@435	2348	return true;
duke@435	2349	}
duke@435	2350
duke@435	2351	// Build field offset expression.
duke@435	2352	// We currently rely on the cookies produced by Unsafe.xxxFieldOffset
duke@435	2353	// to be plain byte offsets, which are also the same as those accepted
duke@435	2354	// by oopDesc::field_base.
duke@435	2355	assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
duke@435	2356	// 32-bit machines ignore the high half of long offsets
duke@435	2357	offset = ConvL2X(offset);
duke@435	2358	Node* adr = make_unsafe_address(base, offset);
duke@435	2359	const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
duke@435	2360
duke@435	2361	// (Unlike inline_unsafe_access, there seems no point in trying
duke@435	2362	// to refine types. Just use the coarse types here.
duke@435	2363	const Type *value_type = Type::get_const_basic_type(type);
duke@435	2364	Compile::AliasType* alias_type = C->alias_type(adr_type);
duke@435	2365	assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
duke@435	2366	int alias_idx = C->get_alias_index(adr_type);
duke@435	2367
duke@435	2368	// Memory-model-wise, a CAS acts like a little synchronized block,
twisti@1040	2369	// so needs barriers on each side. These don't translate into
duke@435	2370	// actual barriers on most machines, but we still need rest of
duke@435	2371	// compiler to respect ordering.
duke@435	2372
duke@435	2373	insert_mem_bar(Op_MemBarRelease);
duke@435	2374	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2375
duke@435	2376	// 4984716: MemBars must be inserted before this
duke@435	2377	// memory node in order to avoid a false
duke@435	2378	// dependency which will confuse the scheduler.
duke@435	2379	Node *mem = memory(alias_idx);
duke@435	2380
duke@435	2381	// For now, we handle only those cases that actually exist: ints,
duke@435	2382	// longs, and Object. Adding others should be straightforward.
duke@435	2383	Node* cas;
duke@435	2384	switch(type) {
duke@435	2385	case T_INT:
duke@435	2386	cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
duke@435	2387	break;
duke@435	2388	case T_LONG:
duke@435	2389	cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
duke@435	2390	break;
duke@435	2391	case T_OBJECT:
coleenp@548	2392	// reference stores need a store barrier.
duke@435	2393	// (They don't if CAS fails, but it isn't worth checking.)
never@1262	2394	pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
coleenp@548	2395	#ifdef _LP64
kvn@598	2396	if (adr->bottom_type()->is_ptr_to_narrowoop()) {
kvn@656	2397	Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
kvn@656	2398	Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
coleenp@548	2399	cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
kvn@656	2400	newval_enc, oldval_enc));
coleenp@548	2401	} else
coleenp@548	2402	#endif
kvn@656	2403	{
kvn@656	2404	cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
kvn@656	2405	}
duke@435	2406	post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
duke@435	2407	break;
duke@435	2408	default:
duke@435	2409	ShouldNotReachHere();
duke@435	2410	break;
duke@435	2411	}
duke@435	2412
duke@435	2413	// SCMemProjNodes represent the memory state of CAS. Their main
duke@435	2414	// role is to prevent CAS nodes from being optimized away when their
duke@435	2415	// results aren't used.
duke@435	2416	Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
duke@435	2417	set_memory(proj, alias_idx);
duke@435	2418
duke@435	2419	// Add the trailing membar surrounding the access
duke@435	2420	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2421	insert_mem_bar(Op_MemBarAcquire);
duke@435	2422
duke@435	2423	push(cas);
duke@435	2424	return true;
duke@435	2425	}
duke@435	2426
duke@435	2427	bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
duke@435	2428	// This is another variant of inline_unsafe_access, differing in
duke@435	2429	// that it always issues store-store ("release") barrier and ensures
duke@435	2430	// store-atomicity (which only matters for "long").
duke@435	2431
duke@435	2432	if (callee()->is_static()) return false; // caller must have the capability!
duke@435	2433
duke@435	2434	#ifndef PRODUCT
duke@435	2435	{
duke@435	2436	ResourceMark rm;
duke@435	2437	// Check the signatures.
duke@435	2438	ciSignature* sig = signature();
duke@435	2439	#ifdef ASSERT
duke@435	2440	BasicType rtype = sig->return_type()->basic_type();
duke@435	2441	assert(rtype == T_VOID, "must return void");
duke@435	2442	assert(sig->count() == 3, "has 3 arguments");
duke@435	2443	assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
duke@435	2444	assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
duke@435	2445	#endif // ASSERT
duke@435	2446	}
duke@435	2447	#endif //PRODUCT
duke@435	2448
duke@435	2449	// number of stack slots per value argument (1 or 2)
duke@435	2450	int type_words = type2size[type];
duke@435	2451
duke@435	2452	C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@435	2453
duke@435	2454	// Argument words: "this" plus oop plus offset plus value;
duke@435	2455	int nargs = 1 + 1 + 2 + type_words;
duke@435	2456
duke@435	2457	// pop arguments: val, offset, base, and receiver
duke@435	2458	debug_only(int saved_sp = _sp);
duke@435	2459	_sp += nargs;
duke@435	2460	Node* val = (type_words == 1) ? pop() : pop_pair();
duke@435	2461	Node *offset = pop_pair();
duke@435	2462	Node *base = pop();
duke@435	2463	Node *receiver = pop();
duke@435	2464	assert(saved_sp == _sp, "must have correct argument count");
duke@435	2465
duke@435	2466	// Null check receiver.
duke@435	2467	_sp += nargs;
duke@435	2468	do_null_check(receiver, T_OBJECT);
duke@435	2469	_sp -= nargs;
duke@435	2470	if (stopped()) {
duke@435	2471	return true;
duke@435	2472	}
duke@435	2473
duke@435	2474	// Build field offset expression.
duke@435	2475	assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
duke@435	2476	// 32-bit machines ignore the high half of long offsets
duke@435	2477	offset = ConvL2X(offset);
duke@435	2478	Node* adr = make_unsafe_address(base, offset);
duke@435	2479	const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
duke@435	2480	const Type *value_type = Type::get_const_basic_type(type);
duke@435	2481	Compile::AliasType* alias_type = C->alias_type(adr_type);
duke@435	2482
duke@435	2483	insert_mem_bar(Op_MemBarRelease);
duke@435	2484	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2485	// Ensure that the store is atomic for longs:
duke@435	2486	bool require_atomic_access = true;
duke@435	2487	Node* store;
duke@435	2488	if (type == T_OBJECT) // reference stores need a store barrier.
never@1260	2489	store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
duke@435	2490	else {
duke@435	2491	store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
duke@435	2492	}
duke@435	2493	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	2494	return true;
duke@435	2495	}
duke@435	2496
duke@435	2497	bool LibraryCallKit::inline_unsafe_allocate() {
duke@435	2498	if (callee()->is_static()) return false; // caller must have the capability!
duke@435	2499	int nargs = 1 + 1;
duke@435	2500	assert(signature()->size() == nargs-1, "alloc has 1 argument");
duke@435	2501	null_check_receiver(callee()); // check then ignore argument(0)
duke@435	2502	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2503	Node* cls = do_null_check(argument(1), T_OBJECT);
duke@435	2504	_sp -= nargs;
duke@435	2505	if (stopped()) return true;
duke@435	2506
duke@435	2507	Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
duke@435	2508	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2509	kls = do_null_check(kls, T_OBJECT);
duke@435	2510	_sp -= nargs;
duke@435	2511	if (stopped()) return true; // argument was like int.class
duke@435	2512
duke@435	2513	// Note: The argument might still be an illegal value like
duke@435	2514	// Serializable.class or Object[].class. The runtime will handle it.
duke@435	2515	// But we must make an explicit check for initialization.
duke@435	2516	Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
duke@435	2517	Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
duke@435	2518	Node* bits = intcon(instanceKlass::fully_initialized);
duke@435	2519	Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
duke@435	2520	// The 'test' is non-zero if we need to take a slow path.
duke@435	2521
duke@435	2522	Node* obj = new_instance(kls, test);
duke@435	2523	push(obj);
duke@435	2524
duke@435	2525	return true;
duke@435	2526	}
duke@435	2527
duke@435	2528	//------------------------inline_native_time_funcs--------------
duke@435	2529	// inline code for System.currentTimeMillis() and System.nanoTime()
duke@435	2530	// these have the same type and signature
duke@435	2531	bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
duke@435	2532	address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
duke@435	2533	CAST_FROM_FN_PTR(address, os::javaTimeMillis);
duke@435	2534	const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
duke@435	2535	const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
duke@435	2536	const TypePtr* no_memory_effects = NULL;
duke@435	2537	Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
duke@435	2538	Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
duke@435	2539	#ifdef ASSERT
duke@435	2540	Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
duke@435	2541	assert(value_top == top(), "second value must be top");
duke@435	2542	#endif
duke@435	2543	push_pair(value);
duke@435	2544	return true;
duke@435	2545	}
duke@435	2546
duke@435	2547	//------------------------inline_native_currentThread------------------
duke@435	2548	bool LibraryCallKit::inline_native_currentThread() {
duke@435	2549	Node* junk = NULL;
duke@435	2550	push(generate_current_thread(junk));
duke@435	2551	return true;
duke@435	2552	}
duke@435	2553
duke@435	2554	//------------------------inline_native_isInterrupted------------------
duke@435	2555	bool LibraryCallKit::inline_native_isInterrupted() {
duke@435	2556	const int nargs = 1+1; // receiver + boolean
duke@435	2557	assert(nargs == arg_size(), "sanity");
duke@435	2558	// Add a fast path to t.isInterrupted(clear_int):
duke@435	2559	// (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
duke@435	2560	// ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
duke@435	2561	// So, in the common case that the interrupt bit is false,
duke@435	2562	// we avoid making a call into the VM. Even if the interrupt bit
duke@435	2563	// is true, if the clear_int argument is false, we avoid the VM call.
duke@435	2564	// However, if the receiver is not currentThread, we must call the VM,
duke@435	2565	// because there must be some locking done around the operation.
duke@435	2566
duke@435	2567	// We only go to the fast case code if we pass two guards.
duke@435	2568	// Paths which do not pass are accumulated in the slow_region.
duke@435	2569	RegionNode* slow_region = new (C, 1) RegionNode(1);
duke@435	2570	record_for_igvn(slow_region);
duke@435	2571	RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
duke@435	2572	PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
duke@435	2573	enum { no_int_result_path = 1,
duke@435	2574	no_clear_result_path = 2,
duke@435	2575	slow_result_path = 3
duke@435	2576	};
duke@435	2577
duke@435	2578	// (a) Receiving thread must be the current thread.
duke@435	2579	Node* rec_thr = argument(0);
duke@435	2580	Node* tls_ptr = NULL;
duke@435	2581	Node* cur_thr = generate_current_thread(tls_ptr);
duke@435	2582	Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
duke@435	2583	Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
duke@435	2584
duke@435	2585	bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
duke@435	2586	if (!known_current_thread)
duke@435	2587	generate_slow_guard(bol_thr, slow_region);
duke@435	2588
duke@435	2589	// (b) Interrupt bit on TLS must be false.
duke@435	2590	Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
duke@435	2591	Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
duke@435	2592	p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
kvn@1222	2593	// Set the control input on the field _interrupted read to prevent it floating up.
kvn@1222	2594	Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
duke@435	2595	Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
duke@435	2596	Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
duke@435	2597
duke@435	2598	IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
duke@435	2599
duke@435	2600	// First fast path: if (!TLS._interrupted) return false;
duke@435	2601	Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
duke@435	2602	result_rgn->init_req(no_int_result_path, false_bit);
duke@435	2603	result_val->init_req(no_int_result_path, intcon(0));
duke@435	2604
duke@435	2605	// drop through to next case
duke@435	2606	set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
duke@435	2607
duke@435	2608	// (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
duke@435	2609	Node* clr_arg = argument(1);
duke@435	2610	Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
duke@435	2611	Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
duke@435	2612	IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
duke@435	2613
duke@435	2614	// Second fast path: ... else if (!clear_int) return true;
duke@435	2615	Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
duke@435	2616	result_rgn->init_req(no_clear_result_path, false_arg);
duke@435	2617	result_val->init_req(no_clear_result_path, intcon(1));
duke@435	2618
duke@435	2619	// drop through to next case
duke@435	2620	set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
duke@435	2621
duke@435	2622	// (d) Otherwise, go to the slow path.
duke@435	2623	slow_region->add_req(control());
duke@435	2624	set_control( _gvn.transform(slow_region) );
duke@435	2625
duke@435	2626	if (stopped()) {
duke@435	2627	// There is no slow path.
duke@435	2628	result_rgn->init_req(slow_result_path, top());
duke@435	2629	result_val->init_req(slow_result_path, top());
duke@435	2630	} else {
duke@435	2631	// non-virtual because it is a private non-static
duke@435	2632	CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
duke@435	2633
duke@435	2634	Node* slow_val = set_results_for_java_call(slow_call);
duke@435	2635	// this->control() comes from set_results_for_java_call
duke@435	2636
duke@435	2637	// If we know that the result of the slow call will be true, tell the optimizer!
duke@435	2638	if (known_current_thread) slow_val = intcon(1);
duke@435	2639
duke@435	2640	Node* fast_io = slow_call->in(TypeFunc::I_O);
duke@435	2641	Node* fast_mem = slow_call->in(TypeFunc::Memory);
duke@435	2642	// These two phis are pre-filled with copies of of the fast IO and Memory
duke@435	2643	Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
duke@435	2644	Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
duke@435	2645
duke@435	2646	result_rgn->init_req(slow_result_path, control());
duke@435	2647	io_phi ->init_req(slow_result_path, i_o());
duke@435	2648	mem_phi ->init_req(slow_result_path, reset_memory());
duke@435	2649	result_val->init_req(slow_result_path, slow_val);
duke@435	2650
duke@435	2651	set_all_memory( _gvn.transform(mem_phi) );
duke@435	2652	set_i_o( _gvn.transform(io_phi) );
duke@435	2653	}
duke@435	2654
duke@435	2655	push_result(result_rgn, result_val);
duke@435	2656	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	2657
duke@435	2658	return true;
duke@435	2659	}
duke@435	2660
duke@435	2661	//---------------------------load_mirror_from_klass----------------------------
duke@435	2662	// Given a klass oop, load its java mirror (a java.lang.Class oop).
duke@435	2663	Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
duke@435	2664	Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
duke@435	2665	return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
duke@435	2666	}
duke@435	2667
duke@435	2668	//-----------------------load_klass_from_mirror_common-------------------------
duke@435	2669	// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
duke@435	2670	// Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
duke@435	2671	// and branch to the given path on the region.
duke@435	2672	// If never_see_null, take an uncommon trap on null, so we can optimistically
duke@435	2673	// compile for the non-null case.
duke@435	2674	// If the region is NULL, force never_see_null = true.
duke@435	2675	Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
duke@435	2676	bool never_see_null,
duke@435	2677	int nargs,
duke@435	2678	RegionNode* region,
duke@435	2679	int null_path,
duke@435	2680	int offset) {
duke@435	2681	if (region == NULL) never_see_null = true;
duke@435	2682	Node* p = basic_plus_adr(mirror, offset);
duke@435	2683	const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
kvn@599	2684	Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
duke@435	2685	_sp += nargs; // any deopt will start just before call to enclosing method
duke@435	2686	Node* null_ctl = top();
duke@435	2687	kls = null_check_oop(kls, &null_ctl, never_see_null);
duke@435	2688	if (region != NULL) {
duke@435	2689	// Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
duke@435	2690	region->init_req(null_path, null_ctl);
duke@435	2691	} else {
duke@435	2692	assert(null_ctl == top(), "no loose ends");
duke@435	2693	}
duke@435	2694	_sp -= nargs;
duke@435	2695	return kls;
duke@435	2696	}
duke@435	2697
duke@435	2698	//--------------------(inline_native_Class_query helpers)---------------------
duke@435	2699	// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
duke@435	2700	// Fall through if (mods & mask) == bits, take the guard otherwise.
duke@435	2701	Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
duke@435	2702	// Branch around if the given klass has the given modifier bit set.
duke@435	2703	// Like generate_guard, adds a new path onto the region.
duke@435	2704	Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
duke@435	2705	Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
duke@435	2706	Node* mask = intcon(modifier_mask);
duke@435	2707	Node* bits = intcon(modifier_bits);
duke@435	2708	Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
duke@435	2709	Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
duke@435	2710	Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
duke@435	2711	return generate_fair_guard(bol, region);
duke@435	2712	}
duke@435	2713	Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
duke@435	2714	return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
duke@435	2715	}
duke@435	2716
duke@435	2717	//-------------------------inline_native_Class_query-------------------
duke@435	2718	bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
duke@435	2719	int nargs = 1+0; // just the Class mirror, in most cases
duke@435	2720	const Type* return_type = TypeInt::BOOL;
duke@435	2721	Node* prim_return_value = top(); // what happens if it's a primitive class?
duke@435	2722	bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
duke@435	2723	bool expect_prim = false; // most of these guys expect to work on refs
duke@435	2724
duke@435	2725	enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
duke@435	2726
duke@435	2727	switch (id) {
duke@435	2728	case vmIntrinsics::_isInstance:
duke@435	2729	nargs = 1+1; // the Class mirror, plus the object getting queried about
duke@435	2730	// nothing is an instance of a primitive type
duke@435	2731	prim_return_value = intcon(0);
duke@435	2732	break;
duke@435	2733	case vmIntrinsics::_getModifiers:
duke@435	2734	prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
duke@435	2735	assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
duke@435	2736	return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
duke@435	2737	break;
duke@435	2738	case vmIntrinsics::_isInterface:
duke@435	2739	prim_return_value = intcon(0);
duke@435	2740	break;
duke@435	2741	case vmIntrinsics::_isArray:
duke@435	2742	prim_return_value = intcon(0);
duke@435	2743	expect_prim = true; // cf. ObjectStreamClass.getClassSignature
duke@435	2744	break;
duke@435	2745	case vmIntrinsics::_isPrimitive:
duke@435	2746	prim_return_value = intcon(1);
duke@435	2747	expect_prim = true; // obviously
duke@435	2748	break;
duke@435	2749	case vmIntrinsics::_getSuperclass:
duke@435	2750	prim_return_value = null();
duke@435	2751	return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
duke@435	2752	break;
duke@435	2753	case vmIntrinsics::_getComponentType:
duke@435	2754	prim_return_value = null();
duke@435	2755	return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
duke@435	2756	break;
duke@435	2757	case vmIntrinsics::_getClassAccessFlags:
duke@435	2758	prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
duke@435	2759	return_type = TypeInt::INT; // not bool! 6297094
duke@435	2760	break;
duke@435	2761	default:
duke@435	2762	ShouldNotReachHere();
duke@435	2763	}
duke@435	2764
duke@435	2765	Node* mirror = argument(0);
duke@435	2766	Node* obj = (nargs <= 1)? top(): argument(1);
duke@435	2767
duke@435	2768	const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
duke@435	2769	if (mirror_con == NULL) return false; // cannot happen?
duke@435	2770
duke@435	2771	#ifndef PRODUCT
duke@435	2772	if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
duke@435	2773	ciType* k = mirror_con->java_mirror_type();
duke@435	2774	if (k) {
duke@435	2775	tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
duke@435	2776	k->print_name();
duke@435	2777	tty->cr();
duke@435	2778	}
duke@435	2779	}
duke@435	2780	#endif
duke@435	2781
duke@435	2782	// Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
duke@435	2783	RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	2784	record_for_igvn(region);
duke@435	2785	PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
duke@435	2786
duke@435	2787	// The mirror will never be null of Reflection.getClassAccessFlags, however
duke@435	2788	// it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
duke@435	2789	// if it is. See bug 4774291.
duke@435	2790
duke@435	2791	// For Reflection.getClassAccessFlags(), the null check occurs in
duke@435	2792	// the wrong place; see inline_unsafe_access(), above, for a similar
duke@435	2793	// situation.
duke@435	2794	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2795	mirror = do_null_check(mirror, T_OBJECT);
duke@435	2796	_sp -= nargs;
duke@435	2797	// If mirror or obj is dead, only null-path is taken.
duke@435	2798	if (stopped()) return true;
duke@435	2799
duke@435	2800	if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
duke@435	2801
duke@435	2802	// Now load the mirror's klass metaobject, and null-check it.
duke@435	2803	// Side-effects region with the control path if the klass is null.
duke@435	2804	Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
duke@435	2805	region, _prim_path);
duke@435	2806	// If kls is null, we have a primitive mirror.
duke@435	2807	phi->init_req(_prim_path, prim_return_value);
duke@435	2808	if (stopped()) { push_result(region, phi); return true; }
duke@435	2809
duke@435	2810	Node* p; // handy temp
duke@435	2811	Node* null_ctl;
duke@435	2812
duke@435	2813	// Now that we have the non-null klass, we can perform the real query.
duke@435	2814	// For constant classes, the query will constant-fold in LoadNode::Value.
duke@435	2815	Node* query_value = top();
duke@435	2816	switch (id) {
duke@435	2817	case vmIntrinsics::_isInstance:
duke@435	2818	// nothing is an instance of a primitive type
duke@435	2819	query_value = gen_instanceof(obj, kls);
duke@435	2820	break;
duke@435	2821
duke@435	2822	case vmIntrinsics::_getModifiers:
duke@435	2823	p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
duke@435	2824	query_value = make_load(NULL, p, TypeInt::INT, T_INT);
duke@435	2825	break;
duke@435	2826
duke@435	2827	case vmIntrinsics::_isInterface:
duke@435	2828	// (To verify this code sequence, check the asserts in JVM_IsInterface.)
duke@435	2829	if (generate_interface_guard(kls, region) != NULL)
duke@435	2830	// A guard was added. If the guard is taken, it was an interface.
duke@435	2831	phi->add_req(intcon(1));
duke@435	2832	// If we fall through, it's a plain class.
duke@435	2833	query_value = intcon(0);
duke@435	2834	break;
duke@435	2835
duke@435	2836	case vmIntrinsics::_isArray:
duke@435	2837	// (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
duke@435	2838	if (generate_array_guard(kls, region) != NULL)
duke@435	2839	// A guard was added. If the guard is taken, it was an array.
duke@435	2840	phi->add_req(intcon(1));
duke@435	2841	// If we fall through, it's a plain class.
duke@435	2842	query_value = intcon(0);
duke@435	2843	break;
duke@435	2844
duke@435	2845	case vmIntrinsics::_isPrimitive:
duke@435	2846	query_value = intcon(0); // "normal" path produces false
duke@435	2847	break;
duke@435	2848
duke@435	2849	case vmIntrinsics::_getSuperclass:
duke@435	2850	// The rules here are somewhat unfortunate, but we can still do better
duke@435	2851	// with random logic than with a JNI call.
duke@435	2852	// Interfaces store null or Object as _super, but must report null.
duke@435	2853	// Arrays store an intermediate super as _super, but must report Object.
duke@435	2854	// Other types can report the actual _super.
duke@435	2855	// (To verify this code sequence, check the asserts in JVM_IsInterface.)
duke@435	2856	if (generate_interface_guard(kls, region) != NULL)
duke@435	2857	// A guard was added. If the guard is taken, it was an interface.
duke@435	2858	phi->add_req(null());
duke@435	2859	if (generate_array_guard(kls, region) != NULL)
duke@435	2860	// A guard was added. If the guard is taken, it was an array.
duke@435	2861	phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
duke@435	2862	// If we fall through, it's a plain class. Get its _super.
duke@435	2863	p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
kvn@599	2864	kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
duke@435	2865	null_ctl = top();
duke@435	2866	kls = null_check_oop(kls, &null_ctl);
duke@435	2867	if (null_ctl != top()) {
duke@435	2868	// If the guard is taken, Object.superClass is null (both klass and mirror).
duke@435	2869	region->add_req(null_ctl);
duke@435	2870	phi ->add_req(null());
duke@435	2871	}
duke@435	2872	if (!stopped()) {
duke@435	2873	query_value = load_mirror_from_klass(kls);
duke@435	2874	}
duke@435	2875	break;
duke@435	2876
duke@435	2877	case vmIntrinsics::_getComponentType:
duke@435	2878	if (generate_array_guard(kls, region) != NULL) {
duke@435	2879	// Be sure to pin the oop load to the guard edge just created:
duke@435	2880	Node* is_array_ctrl = region->in(region->req()-1);
duke@435	2881	Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
duke@435	2882	Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
duke@435	2883	phi->add_req(cmo);
duke@435	2884	}
duke@435	2885	query_value = null(); // non-array case is null
duke@435	2886	break;
duke@435	2887
duke@435	2888	case vmIntrinsics::_getClassAccessFlags:
duke@435	2889	p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
duke@435	2890	query_value = make_load(NULL, p, TypeInt::INT, T_INT);
duke@435	2891	break;
duke@435	2892
duke@435	2893	default:
duke@435	2894	ShouldNotReachHere();
duke@435	2895	}
duke@435	2896
duke@435	2897	// Fall-through is the normal case of a query to a real class.
duke@435	2898	phi->init_req(1, query_value);
duke@435	2899	region->init_req(1, control());
duke@435	2900
duke@435	2901	push_result(region, phi);
duke@435	2902	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	2903
duke@435	2904	return true;
duke@435	2905	}
duke@435	2906
duke@435	2907	//--------------------------inline_native_subtype_check------------------------
duke@435	2908	// This intrinsic takes the JNI calls out of the heart of
duke@435	2909	// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
duke@435	2910	bool LibraryCallKit::inline_native_subtype_check() {
duke@435	2911	int nargs = 1+1; // the Class mirror, plus the other class getting examined
duke@435	2912
duke@435	2913	// Pull both arguments off the stack.
duke@435	2914	Node* args[2]; // two java.lang.Class mirrors: superc, subc
duke@435	2915	args[0] = argument(0);
duke@435	2916	args[1] = argument(1);
duke@435	2917	Node* klasses[2]; // corresponding Klasses: superk, subk
duke@435	2918	klasses[0] = klasses[1] = top();
duke@435	2919
duke@435	2920	enum {
duke@435	2921	// A full decision tree on {superc is prim, subc is prim}:
duke@435	2922	_prim_0_path = 1, // {P,N} => false
duke@435	2923	// {P,P} & superc!=subc => false
duke@435	2924	_prim_same_path, // {P,P} & superc==subc => true
duke@435	2925	_prim_1_path, // {N,P} => false
duke@435	2926	_ref_subtype_path, // {N,N} & subtype check wins => true
duke@435	2927	_both_ref_path, // {N,N} & subtype check loses => false
duke@435	2928	PATH_LIMIT
duke@435	2929	};
duke@435	2930
duke@435	2931	RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	2932	Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
duke@435	2933	record_for_igvn(region);
duke@435	2934
duke@435	2935	const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
duke@435	2936	const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
duke@435	2937	int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
duke@435	2938
duke@435	2939	// First null-check both mirrors and load each mirror's klass metaobject.
duke@435	2940	int which_arg;
duke@435	2941	for (which_arg = 0; which_arg <= 1; which_arg++) {
duke@435	2942	Node* arg = args[which_arg];
duke@435	2943	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2944	arg = do_null_check(arg, T_OBJECT);
duke@435	2945	_sp -= nargs;
duke@435	2946	if (stopped()) break;
duke@435	2947	args[which_arg] = _gvn.transform(arg);
duke@435	2948
duke@435	2949	Node* p = basic_plus_adr(arg, class_klass_offset);
kvn@599	2950	Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
duke@435	2951	klasses[which_arg] = _gvn.transform(kls);
duke@435	2952	}
duke@435	2953
duke@435	2954	// Having loaded both klasses, test each for null.
duke@435	2955	bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
duke@435	2956	for (which_arg = 0; which_arg <= 1; which_arg++) {
duke@435	2957	Node* kls = klasses[which_arg];
duke@435	2958	Node* null_ctl = top();
duke@435	2959	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	2960	kls = null_check_oop(kls, &null_ctl, never_see_null);
duke@435	2961	_sp -= nargs;
duke@435	2962	int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
duke@435	2963	region->init_req(prim_path, null_ctl);
duke@435	2964	if (stopped()) break;
duke@435	2965	klasses[which_arg] = kls;
duke@435	2966	}
duke@435	2967
duke@435	2968	if (!stopped()) {
duke@435	2969	// now we have two reference types, in klasses[0..1]
duke@435	2970	Node* subk = klasses[1]; // the argument to isAssignableFrom
duke@435	2971	Node* superk = klasses[0]; // the receiver
duke@435	2972	region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
duke@435	2973	// now we have a successful reference subtype check
duke@435	2974	region->set_req(_ref_subtype_path, control());
duke@435	2975	}
duke@435	2976
duke@435	2977	// If both operands are primitive (both klasses null), then
duke@435	2978	// we must return true when they are identical primitives.
duke@435	2979	// It is convenient to test this after the first null klass check.
duke@435	2980	set_control(region->in(_prim_0_path)); // go back to first null check
duke@435	2981	if (!stopped()) {
duke@435	2982	// Since superc is primitive, make a guard for the superc==subc case.
duke@435	2983	Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
duke@435	2984	Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
duke@435	2985	generate_guard(bol_eq, region, PROB_FAIR);
duke@435	2986	if (region->req() == PATH_LIMIT+1) {
duke@435	2987	// A guard was added. If the added guard is taken, superc==subc.
duke@435	2988	region->swap_edges(PATH_LIMIT, _prim_same_path);
duke@435	2989	region->del_req(PATH_LIMIT);
duke@435	2990	}
duke@435	2991	region->set_req(_prim_0_path, control()); // Not equal after all.
duke@435	2992	}
duke@435	2993
duke@435	2994	// these are the only paths that produce 'true':
duke@435	2995	phi->set_req(_prim_same_path, intcon(1));
duke@435	2996	phi->set_req(_ref_subtype_path, intcon(1));
duke@435	2997
duke@435	2998	// pull together the cases:
duke@435	2999	assert(region->req() == PATH_LIMIT, "sane region");
duke@435	3000	for (uint i = 1; i < region->req(); i++) {
duke@435	3001	Node* ctl = region->in(i);
duke@435	3002	if (ctl == NULL || ctl == top()) {
duke@435	3003	region->set_req(i, top());
duke@435	3004	phi ->set_req(i, top());
duke@435	3005	} else if (phi->in(i) == NULL) {
duke@435	3006	phi->set_req(i, intcon(0)); // all other paths produce 'false'
duke@435	3007	}
duke@435	3008	}
duke@435	3009
duke@435	3010	set_control(_gvn.transform(region));
duke@435	3011	push(_gvn.transform(phi));
duke@435	3012
duke@435	3013	return true;
duke@435	3014	}
duke@435	3015
duke@435	3016	//---------------------generate_array_guard_common------------------------
duke@435	3017	Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
duke@435	3018	bool obj_array, bool not_array) {
duke@435	3019	// If obj_array/non_array==false/false:
duke@435	3020	// Branch around if the given klass is in fact an array (either obj or prim).
duke@435	3021	// If obj_array/non_array==false/true:
duke@435	3022	// Branch around if the given klass is not an array klass of any kind.
duke@435	3023	// If obj_array/non_array==true/true:
duke@435	3024	// Branch around if the kls is not an oop array (kls is int[], String, etc.)
duke@435	3025	// If obj_array/non_array==true/false:
duke@435	3026	// Branch around if the kls is an oop array (Object[] or subtype)
duke@435	3027	//
duke@435	3028	// Like generate_guard, adds a new path onto the region.
duke@435	3029	jint layout_con = 0;
duke@435	3030	Node* layout_val = get_layout_helper(kls, layout_con);
duke@435	3031	if (layout_val == NULL) {
duke@435	3032	bool query = (obj_array
duke@435	3033	? Klass::layout_helper_is_objArray(layout_con)
duke@435	3034	: Klass::layout_helper_is_javaArray(layout_con));
duke@435	3035	if (query == not_array) {
duke@435	3036	return NULL; // never a branch
duke@435	3037	} else { // always a branch
duke@435	3038	Node* always_branch = control();
duke@435	3039	if (region != NULL)
duke@435	3040	region->add_req(always_branch);
duke@435	3041	set_control(top());
duke@435	3042	return always_branch;
duke@435	3043	}
duke@435	3044	}
duke@435	3045	// Now test the correct condition.
duke@435	3046	jint nval = (obj_array
duke@435	3047	? ((jint)Klass::_lh_array_tag_type_value
duke@435	3048	<< Klass::_lh_array_tag_shift)
duke@435	3049	: Klass::_lh_neutral_value);
duke@435	3050	Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
duke@435	3051	BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
duke@435	3052	// invert the test if we are looking for a non-array
duke@435	3053	if (not_array) btest = BoolTest(btest).negate();
duke@435	3054	Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
duke@435	3055	return generate_fair_guard(bol, region);
duke@435	3056	}
duke@435	3057
duke@435	3058
duke@435	3059	//-----------------------inline_native_newArray--------------------------
duke@435	3060	bool LibraryCallKit::inline_native_newArray() {
duke@435	3061	int nargs = 2;
duke@435	3062	Node* mirror = argument(0);
duke@435	3063	Node* count_val = argument(1);
duke@435	3064
duke@435	3065	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	3066	mirror = do_null_check(mirror, T_OBJECT);
duke@435	3067	_sp -= nargs;
kvn@598	3068	// If mirror or obj is dead, only null-path is taken.
kvn@598	3069	if (stopped()) return true;
duke@435	3070
duke@435	3071	enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
duke@435	3072	RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	3073	PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
duke@435	3074	TypeInstPtr::NOTNULL);
duke@435	3075	PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
duke@435	3076	PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
duke@435	3077	TypePtr::BOTTOM);
duke@435	3078
duke@435	3079	bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
duke@435	3080	Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
duke@435	3081	nargs,
duke@435	3082	result_reg, _slow_path);
duke@435	3083	Node* normal_ctl = control();
duke@435	3084	Node* no_array_ctl = result_reg->in(_slow_path);
duke@435	3085
duke@435	3086	// Generate code for the slow case. We make a call to newArray().
duke@435	3087	set_control(no_array_ctl);
duke@435	3088	if (!stopped()) {
duke@435	3089	// Either the input type is void.class, or else the
duke@435	3090	// array klass has not yet been cached. Either the
duke@435	3091	// ensuing call will throw an exception, or else it
duke@435	3092	// will cache the array klass for next time.
duke@435	3093	PreserveJVMState pjvms(this);
duke@435	3094	CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
duke@435	3095	Node* slow_result = set_results_for_java_call(slow_call);
duke@435	3096	// this->control() comes from set_results_for_java_call
duke@435	3097	result_reg->set_req(_slow_path, control());
duke@435	3098	result_val->set_req(_slow_path, slow_result);
duke@435	3099	result_io ->set_req(_slow_path, i_o());
duke@435	3100	result_mem->set_req(_slow_path, reset_memory());
duke@435	3101	}
duke@435	3102
duke@435	3103	set_control(normal_ctl);
duke@435	3104	if (!stopped()) {
duke@435	3105	// Normal case: The array type has been cached in the java.lang.Class.
duke@435	3106	// The following call works fine even if the array type is polymorphic.
duke@435	3107	// It could be a dynamic mix of int[], boolean[], Object[], etc.
cfang@1165	3108	Node* obj = new_array(klass_node, count_val, nargs);
duke@435	3109	result_reg->init_req(_normal_path, control());
duke@435	3110	result_val->init_req(_normal_path, obj);
duke@435	3111	result_io ->init_req(_normal_path, i_o());
duke@435	3112	result_mem->init_req(_normal_path, reset_memory());
duke@435	3113	}
duke@435	3114
duke@435	3115	// Return the combined state.
duke@435	3116	set_i_o( _gvn.transform(result_io) );
duke@435	3117	set_all_memory( _gvn.transform(result_mem) );
duke@435	3118	push_result(result_reg, result_val);
duke@435	3119	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	3120
duke@435	3121	return true;
duke@435	3122	}
duke@435	3123
duke@435	3124	//----------------------inline_native_getLength--------------------------
duke@435	3125	bool LibraryCallKit::inline_native_getLength() {
duke@435	3126	if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@435	3127
duke@435	3128	int nargs = 1;
duke@435	3129	Node* array = argument(0);
duke@435	3130
duke@435	3131	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	3132	array = do_null_check(array, T_OBJECT);
duke@435	3133	_sp -= nargs;
duke@435	3134
duke@435	3135	// If array is dead, only null-path is taken.
duke@435	3136	if (stopped()) return true;
duke@435	3137
duke@435	3138	// Deoptimize if it is a non-array.
duke@435	3139	Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
duke@435	3140
duke@435	3141	if (non_array != NULL) {
duke@435	3142	PreserveJVMState pjvms(this);
duke@435	3143	set_control(non_array);
duke@435	3144	_sp += nargs; // push the arguments back on the stack
duke@435	3145	uncommon_trap(Deoptimization::Reason_intrinsic,
duke@435	3146	Deoptimization::Action_maybe_recompile);
duke@435	3147	}
duke@435	3148
duke@435	3149	// If control is dead, only non-array-path is taken.
duke@435	3150	if (stopped()) return true;
duke@435	3151
duke@435	3152	// The works fine even if the array type is polymorphic.
duke@435	3153	// It could be a dynamic mix of int[], boolean[], Object[], etc.
duke@435	3154	push( load_array_length(array) );
duke@435	3155
duke@435	3156	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	3157
duke@435	3158	return true;
duke@435	3159	}
duke@435	3160
duke@435	3161	//------------------------inline_array_copyOf----------------------------
duke@435	3162	bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
duke@435	3163	if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
duke@435	3164
duke@435	3165	// Restore the stack and pop off the arguments.
duke@435	3166	int nargs = 3 + (is_copyOfRange? 1: 0);
duke@435	3167	Node* original = argument(0);
duke@435	3168	Node* start = is_copyOfRange? argument(1): intcon(0);
duke@435	3169	Node* end = is_copyOfRange? argument(2): argument(1);
duke@435	3170	Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
duke@435	3171
duke@435	3172	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	3173	array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
duke@435	3174	original = do_null_check(original, T_OBJECT);
duke@435	3175	_sp -= nargs;
duke@435	3176
duke@435	3177	// Check if a null path was taken unconditionally.
duke@435	3178	if (stopped()) return true;
duke@435	3179
duke@435	3180	Node* orig_length = load_array_length(original);
duke@435	3181
duke@435	3182	Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs,
duke@435	3183	NULL, 0);
duke@435	3184	_sp += nargs; // set original stack for use by uncommon_trap
duke@435	3185	klass_node = do_null_check(klass_node, T_OBJECT);
duke@435	3186	_sp -= nargs;
duke@435	3187
duke@435	3188	RegionNode* bailout = new (C, 1) RegionNode(1);
duke@435	3189	record_for_igvn(bailout);
duke@435	3190
duke@435	3191	// Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
duke@435	3192	// Bail out if that is so.
duke@435	3193	Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
duke@435	3194	if (not_objArray != NULL) {
duke@435	3195	// Improve the klass node's type from the new optimistic assumption:
duke@435	3196	ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
duke@435	3197	const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
duke@435	3198	Node* cast = new (C, 2) CastPPNode(klass_node, akls);
duke@435	3199	cast->init_req(0, control());
duke@435	3200	klass_node = _gvn.transform(cast);
duke@435	3201	}
duke@435	3202
duke@435	3203	// Bail out if either start or end is negative.
duke@435	3204	generate_negative_guard(start, bailout, &start);
duke@435	3205	generate_negative_guard(end, bailout, &end);
duke@435	3206
duke@435	3207	Node* length = end;
duke@435	3208	if (_gvn.type(start) != TypeInt::ZERO) {
duke@435	3209	length = _gvn.transform( new (C, 3) SubINode(end, start) );
duke@435	3210	}
duke@435	3211
duke@435	3212	// Bail out if length is negative.
duke@435	3213	// ...Not needed, since the new_array will throw the right exception.
duke@435	3214	//generate_negative_guard(length, bailout, &length);
duke@435	3215
duke@435	3216	if (bailout->req() > 1) {
duke@435	3217	PreserveJVMState pjvms(this);
duke@435	3218	set_control( _gvn.transform(bailout) );
duke@435	3219	_sp += nargs; // push the arguments back on the stack
duke@435	3220	uncommon_trap(Deoptimization::Reason_intrinsic,
duke@435	3221	Deoptimization::Action_maybe_recompile);
duke@435	3222	}
duke@435	3223
duke@435	3224	if (!stopped()) {
duke@435	3225	// How many elements will we copy from the original?
duke@435	3226	// The answer is MinI(orig_length - start, length).
duke@435	3227	Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
duke@435	3228	Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
duke@435	3229
kvn@1268	3230	const bool raw_mem_only = true;
kvn@1268	3231	Node* newcopy = new_array(klass_node, length, nargs, raw_mem_only);
duke@435	3232
duke@435	3233	// Generate a direct call to the right arraycopy function(s).
duke@435	3234	// We know the copy is disjoint but we might not know if the
duke@435	3235	// oop stores need checking.
duke@435	3236	// Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
duke@435	3237	// This will fail a store-check if x contains any non-nulls.
duke@435	3238	bool disjoint_bases = true;
duke@435	3239	bool length_never_negative = true;
duke@435	3240	generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
duke@435	3241	original, start, newcopy, intcon(0), moved,
kvn@1268	3242	disjoint_bases, length_never_negative);
duke@435	3243
duke@435	3244	push(newcopy);
duke@435	3245	}
duke@435	3246
duke@435	3247	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	3248
duke@435	3249	return true;
duke@435	3250	}
duke@435	3251
duke@435	3252
duke@435	3253	//----------------------generate_virtual_guard---------------------------
duke@435	3254	// Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
duke@435	3255	Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
duke@435	3256	RegionNode* slow_region) {
duke@435	3257	ciMethod* method = callee();
duke@435	3258	int vtable_index = method->vtable_index();
duke@435	3259	// Get the methodOop out of the appropriate vtable entry.
duke@435	3260	int entry_offset = (instanceKlass::vtable_start_offset() +
duke@435	3261	vtable_index*vtableEntry::size()) * wordSize +
duke@435	3262	vtableEntry::method_offset_in_bytes();
duke@435	3263	Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
duke@435	3264	Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
duke@435	3265
duke@435	3266	// Compare the target method with the expected method (e.g., Object.hashCode).
duke@435	3267	const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
duke@435	3268
duke@435	3269	Node* native_call = makecon(native_call_addr);
duke@435	3270	Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
duke@435	3271	Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
duke@435	3272
duke@435	3273	return generate_slow_guard(test_native, slow_region);
duke@435	3274	}
duke@435	3275
duke@435	3276	//-----------------------generate_method_call----------------------------
duke@435	3277	// Use generate_method_call to make a slow-call to the real
duke@435	3278	// method if the fast path fails. An alternative would be to
duke@435	3279	// use a stub like OptoRuntime::slow_arraycopy_Java.
duke@435	3280	// This only works for expanding the current library call,
duke@435	3281	// not another intrinsic. (E.g., don't use this for making an
duke@435	3282	// arraycopy call inside of the copyOf intrinsic.)
duke@435	3283	CallJavaNode*
duke@435	3284	LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
duke@435	3285	// When compiling the intrinsic method itself, do not use this technique.
duke@435	3286	guarantee(callee() != C->method(), "cannot make slow-call to self");
duke@435	3287
duke@435	3288	ciMethod* method = callee();
duke@435	3289	// ensure the JVMS we have will be correct for this call
duke@435	3290	guarantee(method_id == method->intrinsic_id(), "must match");
duke@435	3291
duke@435	3292	const TypeFunc* tf = TypeFunc::make(method);
duke@435	3293	int tfdc = tf->domain()->cnt();
duke@435	3294	CallJavaNode* slow_call;
duke@435	3295	if (is_static) {
duke@435	3296	assert(!is_virtual, "");
duke@435	3297	slow_call = new(C, tfdc) CallStaticJavaNode(tf,
duke@435	3298	SharedRuntime::get_resolve_static_call_stub(),
duke@435	3299	method, bci());
duke@435	3300	} else if (is_virtual) {
duke@435	3301	null_check_receiver(method);
duke@435	3302	int vtable_index = methodOopDesc::invalid_vtable_index;
duke@435	3303	if (UseInlineCaches) {
duke@435	3304	// Suppress the vtable call
duke@435	3305	} else {
duke@435	3306	// hashCode and clone are not a miranda methods,
duke@435	3307	// so the vtable index is fixed.
duke@435	3308	// No need to use the linkResolver to get it.
duke@435	3309	vtable_index = method->vtable_index();
duke@435	3310	}
duke@435	3311	slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
duke@435	3312	SharedRuntime::get_resolve_virtual_call_stub(),
duke@435	3313	method, vtable_index, bci());
duke@435	3314	} else { // neither virtual nor static: opt_virtual
duke@435	3315	null_check_receiver(method);
duke@435	3316	slow_call = new(C, tfdc) CallStaticJavaNode(tf,
duke@435	3317	SharedRuntime::get_resolve_opt_virtual_call_stub(),
duke@435	3318	method, bci());
duke@435	3319	slow_call->set_optimized_virtual(true);
duke@435	3320	}
duke@435	3321	set_arguments_for_java_call(slow_call);
duke@435	3322	set_edges_for_java_call(slow_call);
duke@435	3323	return slow_call;
duke@435	3324	}
duke@435	3325
duke@435	3326
duke@435	3327	//------------------------------inline_native_hashcode--------------------
duke@435	3328	// Build special case code for calls to hashCode on an object.
duke@435	3329	bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
duke@435	3330	assert(is_static == callee()->is_static(), "correct intrinsic selection");
duke@435	3331	assert(!(is_virtual && is_static), "either virtual, special, or static");
duke@435	3332
duke@435	3333	enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
duke@435	3334
duke@435	3335	RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	3336	PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
duke@435	3337	TypeInt::INT);
duke@435	3338	PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
duke@435	3339	PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
duke@435	3340	TypePtr::BOTTOM);
duke@435	3341	Node* obj = NULL;
duke@435	3342	if (!is_static) {
duke@435	3343	// Check for hashing null object
duke@435	3344	obj = null_check_receiver(callee());
duke@435	3345	if (stopped()) return true; // unconditionally null
duke@435	3346	result_reg->init_req(_null_path, top());
duke@435	3347	result_val->init_req(_null_path, top());
duke@435	3348	} else {
duke@435	3349	// Do a null check, and return zero if null.
duke@435	3350	// System.identityHashCode(null) == 0
duke@435	3351	obj = argument(0);
duke@435	3352	Node* null_ctl = top();
duke@435	3353	obj = null_check_oop(obj, &null_ctl);
duke@435	3354	result_reg->init_req(_null_path, null_ctl);
duke@435	3355	result_val->init_req(_null_path, _gvn.intcon(0));
duke@435	3356	}
duke@435	3357
duke@435	3358	// Unconditionally null? Then return right away.
duke@435	3359	if (stopped()) {
duke@435	3360	set_control( result_reg->in(_null_path) );
duke@435	3361	if (!stopped())
duke@435	3362	push( result_val ->in(_null_path) );
duke@435	3363	return true;
duke@435	3364	}
duke@435	3365
duke@435	3366	// After null check, get the object's klass.
duke@435	3367	Node* obj_klass = load_object_klass(obj);
duke@435	3368
duke@435	3369	// This call may be virtual (invokevirtual) or bound (invokespecial).
duke@435	3370	// For each case we generate slightly different code.
duke@435	3371
duke@435	3372	// We only go to the fast case code if we pass a number of guards. The
duke@435	3373	// paths which do not pass are accumulated in the slow_region.
duke@435	3374	RegionNode* slow_region = new (C, 1) RegionNode(1);
duke@435	3375	record_for_igvn(slow_region);
duke@435	3376
duke@435	3377	// If this is a virtual call, we generate a funny guard. We pull out
duke@435	3378	// the vtable entry corresponding to hashCode() from the target object.
duke@435	3379	// If the target method which we are calling happens to be the native
duke@435	3380	// Object hashCode() method, we pass the guard. We do not need this
duke@435	3381	// guard for non-virtual calls -- the caller is known to be the native
duke@435	3382	// Object hashCode().
duke@435	3383	if (is_virtual) {
duke@435	3384	generate_virtual_guard(obj_klass, slow_region);
duke@435	3385	}
duke@435	3386
duke@435	3387	// Get the header out of the object, use LoadMarkNode when available
duke@435	3388	Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
duke@435	3389	Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS);
duke@435	3390	header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) );
duke@435	3391
duke@435	3392	// Test the header to see if it is unlocked.
duke@435	3393	Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
duke@435	3394	Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
duke@435	3395	Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
duke@435	3396	Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
duke@435	3397	Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
duke@435	3398
duke@435	3399	generate_slow_guard(test_unlocked, slow_region);
duke@435	3400
duke@435	3401	// Get the hash value and check to see that it has been properly assigned.
duke@435	3402	// We depend on hash_mask being at most 32 bits and avoid the use of
duke@435	3403	// hash_mask_in_place because it could be larger than 32 bits in a 64-bit
duke@435	3404	// vm: see markOop.hpp.
duke@435	3405	Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
duke@435	3406	Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
duke@435	3407	Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
duke@435	3408	// This hack lets the hash bits live anywhere in the mark object now, as long
twisti@1040	3409	// as the shift drops the relevant bits into the low 32 bits. Note that
duke@435	3410	// Java spec says that HashCode is an int so there's no point in capturing
duke@435	3411	// an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
duke@435	3412	hshifted_header = ConvX2I(hshifted_header);
duke@435	3413	Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
duke@435	3414
duke@435	3415	Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
duke@435	3416	Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
duke@435	3417	Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
duke@435	3418
duke@435	3419	generate_slow_guard(test_assigned, slow_region);
duke@435	3420
duke@435	3421	Node* init_mem = reset_memory();
duke@435	3422	// fill in the rest of the null path:
duke@435	3423	result_io ->init_req(_null_path, i_o());
duke@435	3424	result_mem->init_req(_null_path, init_mem);
duke@435	3425
duke@435	3426	result_val->init_req(_fast_path, hash_val);
duke@435	3427	result_reg->init_req(_fast_path, control());
duke@435	3428	result_io ->init_req(_fast_path, i_o());
duke@435	3429	result_mem->init_req(_fast_path, init_mem);
duke@435	3430
duke@435	3431	// Generate code for the slow case. We make a call to hashCode().
duke@435	3432	set_control(_gvn.transform(slow_region));
duke@435	3433	if (!stopped()) {
duke@435	3434	// No need for PreserveJVMState, because we're using up the present state.
duke@435	3435	set_all_memory(init_mem);
duke@435	3436	vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
duke@435	3437	if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
duke@435	3438	CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
duke@435	3439	Node* slow_result = set_results_for_java_call(slow_call);
duke@435	3440	// this->control() comes from set_results_for_java_call
duke@435	3441	result_reg->init_req(_slow_path, control());
duke@435	3442	result_val->init_req(_slow_path, slow_result);
duke@435	3443	result_io ->set_req(_slow_path, i_o());
duke@435	3444	result_mem ->set_req(_slow_path, reset_memory());
duke@435	3445	}
duke@435	3446
duke@435	3447	// Return the combined state.
duke@435	3448	set_i_o( _gvn.transform(result_io) );
duke@435	3449	set_all_memory( _gvn.transform(result_mem) );
duke@435	3450	push_result(result_reg, result_val);
duke@435	3451
duke@435	3452	return true;
duke@435	3453	}
duke@435	3454
duke@435	3455	//---------------------------inline_native_getClass----------------------------
twisti@1040	3456	// Build special case code for calls to getClass on an object.
duke@435	3457	bool LibraryCallKit::inline_native_getClass() {
duke@435	3458	Node* obj = null_check_receiver(callee());
duke@435	3459	if (stopped()) return true;
duke@435	3460	push( load_mirror_from_klass(load_object_klass(obj)) );
duke@435	3461	return true;
duke@435	3462	}
duke@435	3463
duke@435	3464	//-----------------inline_native_Reflection_getCallerClass---------------------
duke@435	3465	// In the presence of deep enough inlining, getCallerClass() becomes a no-op.
duke@435	3466	//
duke@435	3467	// NOTE that this code must perform the same logic as
duke@435	3468	// vframeStream::security_get_caller_frame in that it must skip
duke@435	3469	// Method.invoke() and auxiliary frames.
duke@435	3470
duke@435	3471
duke@435	3472
duke@435	3473
duke@435	3474	bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
duke@435	3475	ciMethod* method = callee();
duke@435	3476
duke@435	3477	#ifndef PRODUCT
duke@435	3478	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3479	tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
duke@435	3480	}
duke@435	3481	#endif
duke@435	3482
duke@435	3483	debug_only(int saved_sp = _sp);
duke@435	3484
duke@435	3485	// Argument words: (int depth)
duke@435	3486	int nargs = 1;
duke@435	3487
duke@435	3488	_sp += nargs;
duke@435	3489	Node* caller_depth_node = pop();
duke@435	3490
duke@435	3491	assert(saved_sp == _sp, "must have correct argument count");
duke@435	3492
duke@435	3493	// The depth value must be a constant in order for the runtime call
duke@435	3494	// to be eliminated.
duke@435	3495	const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
duke@435	3496	if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
duke@435	3497	#ifndef PRODUCT
duke@435	3498	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3499	tty->print_cr(" Bailing out because caller depth was not a constant");
duke@435	3500	}
duke@435	3501	#endif
duke@435	3502	return false;
duke@435	3503	}
duke@435	3504	// Note that the JVM state at this point does not include the
duke@435	3505	// getCallerClass() frame which we are trying to inline. The
duke@435	3506	// semantics of getCallerClass(), however, are that the "first"
duke@435	3507	// frame is the getCallerClass() frame, so we subtract one from the
duke@435	3508	// requested depth before continuing. We don't inline requests of
duke@435	3509	// getCallerClass(0).
duke@435	3510	int caller_depth = caller_depth_type->get_con() - 1;
duke@435	3511	if (caller_depth < 0) {
duke@435	3512	#ifndef PRODUCT
duke@435	3513	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3514	tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
duke@435	3515	}
duke@435	3516	#endif
duke@435	3517	return false;
duke@435	3518	}
duke@435	3519
duke@435	3520	if (!jvms()->has_method()) {
duke@435	3521	#ifndef PRODUCT
duke@435	3522	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3523	tty->print_cr(" Bailing out because intrinsic was inlined at top level");
duke@435	3524	}
duke@435	3525	#endif
duke@435	3526	return false;
duke@435	3527	}
duke@435	3528	int _depth = jvms()->depth(); // cache call chain depth
duke@435	3529
duke@435	3530	// Walk back up the JVM state to find the caller at the required
duke@435	3531	// depth. NOTE that this code must perform the same logic as
duke@435	3532	// vframeStream::security_get_caller_frame in that it must skip
duke@435	3533	// Method.invoke() and auxiliary frames. Note also that depth is
duke@435	3534	// 1-based (1 is the bottom of the inlining).
duke@435	3535	int inlining_depth = _depth;
duke@435	3536	JVMState* caller_jvms = NULL;
duke@435	3537
duke@435	3538	if (inlining_depth > 0) {
duke@435	3539	caller_jvms = jvms();
duke@435	3540	assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
duke@435	3541	do {
duke@435	3542	// The following if-tests should be performed in this order
duke@435	3543	if (is_method_invoke_or_aux_frame(caller_jvms)) {
duke@435	3544	// Skip a Method.invoke() or auxiliary frame
duke@435	3545	} else if (caller_depth > 0) {
duke@435	3546	// Skip real frame
duke@435	3547	--caller_depth;
duke@435	3548	} else {
duke@435	3549	// We're done: reached desired caller after skipping.
duke@435	3550	break;
duke@435	3551	}
duke@435	3552	caller_jvms = caller_jvms->caller();
duke@435	3553	--inlining_depth;
duke@435	3554	} while (inlining_depth > 0);
duke@435	3555	}
duke@435	3556
duke@435	3557	if (inlining_depth == 0) {
duke@435	3558	#ifndef PRODUCT
duke@435	3559	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3560	tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
duke@435	3561	tty->print_cr(" JVM state at this point:");
duke@435	3562	for (int i = _depth; i >= 1; i--) {
duke@435	3563	tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
duke@435	3564	}
duke@435	3565	}
duke@435	3566	#endif
duke@435	3567	return false; // Reached end of inlining
duke@435	3568	}
duke@435	3569
duke@435	3570	// Acquire method holder as java.lang.Class
duke@435	3571	ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
duke@435	3572	ciInstance* caller_mirror = caller_klass->java_mirror();
duke@435	3573	// Push this as a constant
duke@435	3574	push(makecon(TypeInstPtr::make(caller_mirror)));
duke@435	3575	#ifndef PRODUCT
duke@435	3576	if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
duke@435	3577	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	3578	tty->print_cr(" JVM state at this point:");
duke@435	3579	for (int i = _depth; i >= 1; i--) {
duke@435	3580	tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
duke@435	3581	}
duke@435	3582	}
duke@435	3583	#endif
duke@435	3584	return true;
duke@435	3585	}
duke@435	3586
duke@435	3587	// Helper routine for above
duke@435	3588	bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
duke@435	3589	// Is this the Method.invoke method itself?
duke@435	3590	if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke)
duke@435	3591	return true;
duke@435	3592
duke@435	3593	// Is this a helper, defined somewhere underneath MethodAccessorImpl.
duke@435	3594	ciKlass* k = jvms->method()->holder();
duke@435	3595	if (k->is_instance_klass()) {
duke@435	3596	ciInstanceKlass* ik = k->as_instance_klass();
duke@435	3597	for (; ik != NULL; ik = ik->super()) {
duke@435	3598	if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
duke@435	3599	ik == env()->find_system_klass(ik->name())) {
duke@435	3600	return true;
duke@435	3601	}
duke@435	3602	}
duke@435	3603	}
duke@435	3604
duke@435	3605	return false;
duke@435	3606	}
duke@435	3607
duke@435	3608	static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
duke@435	3609	// inline_native_AtomicLong_attemptUpdate() but it has no way of
duke@435	3610	// computing it since there is no lookup field by name function in the
duke@435	3611	// CI interface. This is computed and set by inline_native_AtomicLong_get().
duke@435	3612	// Using a static variable here is safe even if we have multiple compilation
duke@435	3613	// threads because the offset is constant. At worst the same offset will be
duke@435	3614	// computed and stored multiple
duke@435	3615
duke@435	3616	bool LibraryCallKit::inline_native_AtomicLong_get() {
duke@435	3617	// Restore the stack and pop off the argument
duke@435	3618	_sp+=1;
duke@435	3619	Node *obj = pop();
duke@435	3620
duke@435	3621	// get the offset of the "value" field. Since the CI interfaces
duke@435	3622	// does not provide a way to look up a field by name, we scan the bytecodes
duke@435	3623	// to get the field index. We expect the first 2 instructions of the method
duke@435	3624	// to be:
duke@435	3625	// 0 aload_0
duke@435	3626	// 1 getfield "value"
duke@435	3627	ciMethod* method = callee();
duke@435	3628	if (value_field_offset == -1)
duke@435	3629	{
duke@435	3630	ciField* value_field;
duke@435	3631	ciBytecodeStream iter(method);
duke@435	3632	Bytecodes::Code bc = iter.next();
duke@435	3633
duke@435	3634	if ((bc != Bytecodes::_aload_0) &&
duke@435	3635	((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
duke@435	3636	return false;
duke@435	3637	bc = iter.next();
duke@435	3638	if (bc != Bytecodes::_getfield)
duke@435	3639	return false;
duke@435	3640	bool ignore;
duke@435	3641	value_field = iter.get_field(ignore);
duke@435	3642	value_field_offset = value_field->offset_in_bytes();
duke@435	3643	}
duke@435	3644
duke@435	3645	// Null check without removing any arguments.
duke@435	3646	_sp++;
duke@435	3647	obj = do_null_check(obj, T_OBJECT);
duke@435	3648	_sp--;
duke@435	3649	// Check for locking null object
duke@435	3650	if (stopped()) return true;
duke@435	3651
duke@435	3652	Node *adr = basic_plus_adr(obj, obj, value_field_offset);
duke@435	3653	const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
duke@435	3654	int alias_idx = C->get_alias_index(adr_type);
duke@435	3655
duke@435	3656	Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
duke@435	3657
duke@435	3658	push_pair(result);
duke@435	3659
duke@435	3660	return true;
duke@435	3661	}
duke@435	3662
duke@435	3663	bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
duke@435	3664	// Restore the stack and pop off the arguments
duke@435	3665	_sp+=5;
duke@435	3666	Node *newVal = pop_pair();
duke@435	3667	Node *oldVal = pop_pair();
duke@435	3668	Node *obj = pop();
duke@435	3669
duke@435	3670	// we need the offset of the "value" field which was computed when
duke@435	3671	// inlining the get() method. Give up if we don't have it.
duke@435	3672	if (value_field_offset == -1)
duke@435	3673	return false;
duke@435	3674
duke@435	3675	// Null check without removing any arguments.
duke@435	3676	_sp+=5;
duke@435	3677	obj = do_null_check(obj, T_OBJECT);
duke@435	3678	_sp-=5;
duke@435	3679	// Check for locking null object
duke@435	3680	if (stopped()) return true;
duke@435	3681
duke@435	3682	Node *adr = basic_plus_adr(obj, obj, value_field_offset);
duke@435	3683	const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
duke@435	3684	int alias_idx = C->get_alias_index(adr_type);
duke@435	3685
kvn@855	3686	Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
kvn@855	3687	Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
duke@435	3688	set_memory(store_proj, alias_idx);
kvn@855	3689	Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
kvn@855	3690
kvn@855	3691	Node *result;
kvn@855	3692	// CMove node is not used to be able fold a possible check code
kvn@855	3693	// after attemptUpdate() call. This code could be transformed
kvn@855	3694	// into CMove node by loop optimizations.
kvn@855	3695	{
kvn@855	3696	RegionNode *r = new (C, 3) RegionNode(3);
kvn@855	3697	result = new (C, 3) PhiNode(r, TypeInt::BOOL);
kvn@855	3698
kvn@855	3699	Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
kvn@855	3700	Node *iftrue = opt_iff(r, iff);
kvn@855	3701	r->init_req(1, iftrue);
kvn@855	3702	result->init_req(1, intcon(1));
kvn@855	3703	result->init_req(2, intcon(0));
kvn@855	3704
kvn@855	3705	set_control(_gvn.transform(r));
kvn@855	3706	record_for_igvn(r);
kvn@855	3707
kvn@855	3708	C->set_has_split_ifs(true); // Has chance for split-if optimization
kvn@855	3709	}
kvn@855	3710
kvn@855	3711	push(_gvn.transform(result));
duke@435	3712	return true;
duke@435	3713	}
duke@435	3714
duke@435	3715	bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
duke@435	3716	// restore the arguments
duke@435	3717	_sp += arg_size();
duke@435	3718
duke@435	3719	switch (id) {
duke@435	3720	case vmIntrinsics::_floatToRawIntBits:
duke@435	3721	push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
duke@435	3722	break;
duke@435	3723
duke@435	3724	case vmIntrinsics::_intBitsToFloat:
duke@435	3725	push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
duke@435	3726	break;
duke@435	3727
duke@435	3728	case vmIntrinsics::_doubleToRawLongBits:
duke@435	3729	push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
duke@435	3730	break;
duke@435	3731
duke@435	3732	case vmIntrinsics::_longBitsToDouble:
duke@435	3733	push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
duke@435	3734	break;
duke@435	3735
duke@435	3736	case vmIntrinsics::_doubleToLongBits: {
duke@435	3737	Node* value = pop_pair();
duke@435	3738
duke@435	3739	// two paths (plus control) merge in a wood
duke@435	3740	RegionNode *r = new (C, 3) RegionNode(3);
duke@435	3741	Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
duke@435	3742
duke@435	3743	Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
duke@435	3744	// Build the boolean node
duke@435	3745	Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
duke@435	3746
duke@435	3747	// Branch either way.
duke@435	3748	// NaN case is less traveled, which makes all the difference.
duke@435	3749	IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
duke@435	3750	Node *opt_isnan = _gvn.transform(ifisnan);
duke@435	3751	assert( opt_isnan->is_If(), "Expect an IfNode");
duke@435	3752	IfNode *opt_ifisnan = (IfNode*)opt_isnan;
duke@435	3753	Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
duke@435	3754
duke@435	3755	set_control(iftrue);
duke@435	3756
duke@435	3757	static const jlong nan_bits = CONST64(0x7ff8000000000000);
duke@435	3758	Node *slow_result = longcon(nan_bits); // return NaN
duke@435	3759	phi->init_req(1, _gvn.transform( slow_result ));
duke@435	3760	r->init_req(1, iftrue);
duke@435	3761
duke@435	3762	// Else fall through
duke@435	3763	Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
duke@435	3764	set_control(iffalse);
duke@435	3765
duke@435	3766	phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
duke@435	3767	r->init_req(2, iffalse);
duke@435	3768
duke@435	3769	// Post merge
duke@435	3770	set_control(_gvn.transform(r));
duke@435	3771	record_for_igvn(r);
duke@435	3772
duke@435	3773	Node* result = _gvn.transform(phi);
duke@435	3774	assert(result->bottom_type()->isa_long(), "must be");
duke@435	3775	push_pair(result);
duke@435	3776
duke@435	3777	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	3778
duke@435	3779	break;
duke@435	3780	}
duke@435	3781
duke@435	3782	case vmIntrinsics::_floatToIntBits: {
duke@435	3783	Node* value = pop();
duke@435	3784
duke@435	3785	// two paths (plus control) merge in a wood
duke@435	3786	RegionNode *r = new (C, 3) RegionNode(3);
duke@435	3787	Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
duke@435	3788
duke@435	3789	Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
duke@435	3790	// Build the boolean node
duke@435	3791	Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
duke@435	3792
duke@435	3793	// Branch either way.
duke@435	3794	// NaN case is less traveled, which makes all the difference.
duke@435	3795	IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
duke@435	3796	Node *opt_isnan = _gvn.transform(ifisnan);
duke@435	3797	assert( opt_isnan->is_If(), "Expect an IfNode");
duke@435	3798	IfNode *opt_ifisnan = (IfNode*)opt_isnan;
duke@435	3799	Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
duke@435	3800
duke@435	3801	set_control(iftrue);
duke@435	3802
duke@435	3803	static const jint nan_bits = 0x7fc00000;
duke@435	3804	Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
duke@435	3805	phi->init_req(1, _gvn.transform( slow_result ));
duke@435	3806	r->init_req(1, iftrue);
duke@435	3807
duke@435	3808	// Else fall through
duke@435	3809	Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
duke@435	3810	set_control(iffalse);
duke@435	3811
duke@435	3812	phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
duke@435	3813	r->init_req(2, iffalse);
duke@435	3814
duke@435	3815	// Post merge
duke@435	3816	set_control(_gvn.transform(r));
duke@435	3817	record_for_igvn(r);
duke@435	3818
duke@435	3819	Node* result = _gvn.transform(phi);
duke@435	3820	assert(result->bottom_type()->isa_int(), "must be");
duke@435	3821	push(result);
duke@435	3822
duke@435	3823	C->set_has_split_ifs(true); // Has chance for split-if optimization
duke@435	3824
duke@435	3825	break;
duke@435	3826	}
duke@435	3827
duke@435	3828	default:
duke@435	3829	ShouldNotReachHere();
duke@435	3830	}
duke@435	3831
duke@435	3832	return true;
duke@435	3833	}
duke@435	3834
duke@435	3835	#ifdef _LP64
duke@435	3836	#define XTOP ,top() /*additional argument*/
duke@435	3837	#else //_LP64
duke@435	3838	#define XTOP /*no additional argument*/
duke@435	3839	#endif //_LP64
duke@435	3840
duke@435	3841	//----------------------inline_unsafe_copyMemory-------------------------
duke@435	3842	bool LibraryCallKit::inline_unsafe_copyMemory() {
duke@435	3843	if (callee()->is_static()) return false; // caller must have the capability!
duke@435	3844	int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
duke@435	3845	assert(signature()->size() == nargs-1, "copy has 5 arguments");
duke@435	3846	null_check_receiver(callee()); // check then ignore argument(0)
duke@435	3847	if (stopped()) return true;
duke@435	3848
duke@435	3849	C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
duke@435	3850
duke@435	3851	Node* src_ptr = argument(1);
duke@435	3852	Node* src_off = ConvL2X(argument(2));
duke@435	3853	assert(argument(3)->is_top(), "2nd half of long");
duke@435	3854	Node* dst_ptr = argument(4);
duke@435	3855	Node* dst_off = ConvL2X(argument(5));
duke@435	3856	assert(argument(6)->is_top(), "2nd half of long");
duke@435	3857	Node* size = ConvL2X(argument(7));
duke@435	3858	assert(argument(8)->is_top(), "2nd half of long");
duke@435	3859
duke@435	3860	assert(Unsafe_field_offset_to_byte_offset(11) == 11,
duke@435	3861	"fieldOffset must be byte-scaled");
duke@435	3862
duke@435	3863	Node* src = make_unsafe_address(src_ptr, src_off);
duke@435	3864	Node* dst = make_unsafe_address(dst_ptr, dst_off);
duke@435	3865
duke@435	3866	// Conservatively insert a memory barrier on all memory slices.
duke@435	3867	// Do not let writes of the copy source or destination float below the copy.
duke@435	3868	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	3869
duke@435	3870	// Call it. Note that the length argument is not scaled.
duke@435	3871	make_runtime_call(RC_LEAF|RC_NO_FP,
duke@435	3872	OptoRuntime::fast_arraycopy_Type(),
duke@435	3873	StubRoutines::unsafe_arraycopy(),
duke@435	3874	"unsafe_arraycopy",
duke@435	3875	TypeRawPtr::BOTTOM,
duke@435	3876	src, dst, size XTOP);
duke@435	3877
duke@435	3878	// Do not let reads of the copy destination float above the copy.
duke@435	3879	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	3880
duke@435	3881	return true;
duke@435	3882	}
duke@435	3883
kvn@1268	3884	//------------------------clone_coping-----------------------------------
kvn@1268	3885	// Helper function for inline_native_clone.
kvn@1268	3886	void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
kvn@1268	3887	assert(obj_size != NULL, "");
kvn@1268	3888	Node* raw_obj = alloc_obj->in(1);
kvn@1268	3889	assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
kvn@1268	3890
kvn@1268	3891	if (ReduceBulkZeroing) {
kvn@1268	3892	// We will be completely responsible for initializing this object -
kvn@1268	3893	// mark Initialize node as complete.
kvn@1268	3894	AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
kvn@1268	3895	// The object was just allocated - there should be no any stores!
kvn@1268	3896	guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
kvn@1268	3897	}
kvn@1268	3898
kvn@1268	3899	// Cast to Object for arraycopy.
kvn@1268	3900	// We can't use the original CheckCastPP since it should be moved
kvn@1268	3901	// after the arraycopy to prevent stores flowing above it.
kvn@1268	3902	Node* new_obj = new(C, 2) CheckCastPPNode(alloc_obj->in(0), raw_obj,
kvn@1268	3903	TypeInstPtr::NOTNULL);
kvn@1268	3904	new_obj = _gvn.transform(new_obj);
kvn@1268	3905	// Substitute in the locally valid dest_oop.
kvn@1268	3906	replace_in_map(alloc_obj, new_obj);
kvn@1268	3907
kvn@1268	3908	// Copy the fastest available way.
kvn@1268	3909	// TODO: generate fields copies for small objects instead.
kvn@1268	3910	Node* src = obj;
kvn@1268	3911	Node* dest = new_obj;
kvn@1268	3912	Node* size = _gvn.transform(obj_size);
kvn@1268	3913
kvn@1268	3914	// Exclude the header but include array length to copy by 8 bytes words.
kvn@1268	3915	// Can't use base_offset_in_bytes(bt) since basic type is unknown.
kvn@1268	3916	int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
kvn@1268	3917	instanceOopDesc::base_offset_in_bytes();
kvn@1268	3918	// base_off:
kvn@1268	3919	// 8 - 32-bit VM
kvn@1268	3920	// 12 - 64-bit VM, compressed oops
kvn@1268	3921	// 16 - 64-bit VM, normal oops
kvn@1268	3922	if (base_off % BytesPerLong != 0) {
kvn@1268	3923	assert(UseCompressedOops, "");
kvn@1268	3924	if (is_array) {
kvn@1268	3925	// Exclude length to copy by 8 bytes words.
kvn@1268	3926	base_off += sizeof(int);
kvn@1268	3927	} else {
kvn@1268	3928	// Include klass to copy by 8 bytes words.
kvn@1268	3929	base_off = instanceOopDesc::klass_offset_in_bytes();
kvn@1268	3930	}
kvn@1268	3931	assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
kvn@1268	3932	}
kvn@1268	3933	src = basic_plus_adr(src, base_off);
kvn@1268	3934	dest = basic_plus_adr(dest, base_off);
kvn@1268	3935
kvn@1268	3936	// Compute the length also, if needed:
kvn@1268	3937	Node* countx = size;
kvn@1268	3938	countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
kvn@1268	3939	countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
kvn@1268	3940
kvn@1268	3941	const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
kvn@1268	3942	bool disjoint_bases = true;
kvn@1268	3943	generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
kvn@1268	3944	src, NULL, dest, NULL, countx);
kvn@1268	3945
kvn@1268	3946	// If necessary, emit some card marks afterwards. (Non-arrays only.)
kvn@1268	3947	if (card_mark) {
kvn@1268	3948	assert(!is_array, "");
kvn@1268	3949	// Put in store barrier for any and all oops we are sticking
kvn@1268	3950	// into this object. (We could avoid this if we could prove
kvn@1268	3951	// that the object type contains no oop fields at all.)
kvn@1268	3952	Node* no_particular_value = NULL;
kvn@1268	3953	Node* no_particular_field = NULL;
kvn@1268	3954	int raw_adr_idx = Compile::AliasIdxRaw;
kvn@1268	3955	post_barrier(control(),
kvn@1268	3956	memory(raw_adr_type),
kvn@1268	3957	new_obj,
kvn@1268	3958	no_particular_field,
kvn@1268	3959	raw_adr_idx,
kvn@1268	3960	no_particular_value,
kvn@1268	3961	T_OBJECT,
kvn@1268	3962	false);
kvn@1268	3963	}
kvn@1268	3964
kvn@1268	3965	// Move the original CheckCastPP after arraycopy.
kvn@1268	3966	_gvn.hash_delete(alloc_obj);
kvn@1268	3967	alloc_obj->set_req(0, control());
kvn@1268	3968	// Replace raw memory edge with new CheckCastPP to have a live oop
kvn@1268	3969	// at safepoints instead of raw value.
kvn@1268	3970	assert(new_obj->is_CheckCastPP() && new_obj->in(1) == alloc_obj->in(1), "sanity");
kvn@1268	3971	alloc_obj->set_req(1, new_obj); // cast to the original type
kvn@1268	3972	_gvn.hash_find_insert(alloc_obj); // put back into GVN table
kvn@1268	3973	// Restore in the locally valid dest_oop.
kvn@1268	3974	replace_in_map(new_obj, alloc_obj);
kvn@1268	3975	}
duke@435	3976
duke@435	3977	//------------------------inline_native_clone----------------------------
duke@435	3978	// Here are the simple edge cases:
duke@435	3979	// null receiver => normal trap
duke@435	3980	// virtual and clone was overridden => slow path to out-of-line clone
duke@435	3981	// not cloneable or finalizer => slow path to out-of-line Object.clone
duke@435	3982	//
duke@435	3983	// The general case has two steps, allocation and copying.
duke@435	3984	// Allocation has two cases, and uses GraphKit::new_instance or new_array.
duke@435	3985	//
duke@435	3986	// Copying also has two cases, oop arrays and everything else.
duke@435	3987	// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
duke@435	3988	// Everything else uses the tight inline loop supplied by CopyArrayNode.
duke@435	3989	//
duke@435	3990	// These steps fold up nicely if and when the cloned object's klass
duke@435	3991	// can be sharply typed as an object array, a type array, or an instance.
duke@435	3992	//
duke@435	3993	bool LibraryCallKit::inline_native_clone(bool is_virtual) {
duke@435	3994	int nargs = 1;
duke@435	3995	Node* obj = null_check_receiver(callee());
duke@435	3996	if (stopped()) return true;
duke@435	3997	Node* obj_klass = load_object_klass(obj);
duke@435	3998	const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
duke@435	3999	const TypeOopPtr* toop = ((tklass != NULL)
duke@435	4000	? tklass->as_instance_type()
duke@435	4001	: TypeInstPtr::NOTNULL);
duke@435	4002
duke@435	4003	// Conservatively insert a memory barrier on all memory slices.
duke@435	4004	// Do not let writes into the original float below the clone.
duke@435	4005	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	4006
duke@435	4007	// paths into result_reg:
duke@435	4008	enum {
duke@435	4009	_slow_path = 1, // out-of-line call to clone method (virtual or not)
kvn@1268	4010	_objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
kvn@1268	4011	_array_path, // plain array allocation, plus arrayof_long_arraycopy
kvn@1268	4012	_instance_path, // plain instance allocation, plus arrayof_long_arraycopy
duke@435	4013	PATH_LIMIT
duke@435	4014	};
duke@435	4015	RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	4016	PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
duke@435	4017	TypeInstPtr::NOTNULL);
duke@435	4018	PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
duke@435	4019	PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
duke@435	4020	TypePtr::BOTTOM);
duke@435	4021	record_for_igvn(result_reg);
duke@435	4022
duke@435	4023	const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
duke@435	4024	int raw_adr_idx = Compile::AliasIdxRaw;
duke@435	4025	const bool raw_mem_only = true;
duke@435	4026
duke@435	4027	Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
duke@435	4028	if (array_ctl != NULL) {
duke@435	4029	// It's an array.
duke@435	4030	PreserveJVMState pjvms(this);
duke@435	4031	set_control(array_ctl);
duke@435	4032	Node* obj_length = load_array_length(obj);
duke@435	4033	Node* obj_size = NULL;
cfang@1165	4034	Node* alloc_obj = new_array(obj_klass, obj_length, nargs,
duke@435	4035	raw_mem_only, &obj_size);
duke@435	4036
duke@435	4037	if (!use_ReduceInitialCardMarks()) {
duke@435	4038	// If it is an oop array, it requires very special treatment,
duke@435	4039	// because card marking is required on each card of the array.
duke@435	4040	Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
duke@435	4041	if (is_obja != NULL) {
duke@435	4042	PreserveJVMState pjvms2(this);
duke@435	4043	set_control(is_obja);
duke@435	4044	// Generate a direct call to the right arraycopy function(s).
duke@435	4045	bool disjoint_bases = true;
duke@435	4046	bool length_never_negative = true;
duke@435	4047	generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
duke@435	4048	obj, intcon(0), alloc_obj, intcon(0),
kvn@1268	4049	obj_length,
duke@435	4050	disjoint_bases, length_never_negative);
duke@435	4051	result_reg->init_req(_objArray_path, control());
duke@435	4052	result_val->init_req(_objArray_path, alloc_obj);
duke@435	4053	result_i_o ->set_req(_objArray_path, i_o());
duke@435	4054	result_mem ->set_req(_objArray_path, reset_memory());
duke@435	4055	}
duke@435	4056	}
duke@435	4057	// We can dispense with card marks if we know the allocation
duke@435	4058	// comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
duke@435	4059	// causes the non-eden paths to simulate a fresh allocation,
duke@435	4060	// insofar that no further card marks are required to initialize
duke@435	4061	// the object.
duke@435	4062
duke@435	4063	// Otherwise, there are no card marks to worry about.
kvn@1268	4064
kvn@1268	4065	if (!stopped()) {
kvn@1268	4066	copy_to_clone(obj, alloc_obj, obj_size, true, false);
kvn@1268	4067
kvn@1268	4068	// Present the results of the copy.
kvn@1268	4069	result_reg->init_req(_array_path, control());
kvn@1268	4070	result_val->init_req(_array_path, alloc_obj);
kvn@1268	4071	result_i_o ->set_req(_array_path, i_o());
kvn@1268	4072	result_mem ->set_req(_array_path, reset_memory());
kvn@1268	4073	}
duke@435	4074	}
duke@435	4075
kvn@1268	4076	// We only go to the instance fast case code if we pass a number of guards.
duke@435	4077	// The paths which do not pass are accumulated in the slow_region.
duke@435	4078	RegionNode* slow_region = new (C, 1) RegionNode(1);
duke@435	4079	record_for_igvn(slow_region);
duke@435	4080	if (!stopped()) {
kvn@1268	4081	// It's an instance (we did array above). Make the slow-path tests.
duke@435	4082	// If this is a virtual call, we generate a funny guard. We grab
duke@435	4083	// the vtable entry corresponding to clone() from the target object.
duke@435	4084	// If the target method which we are calling happens to be the
duke@435	4085	// Object clone() method, we pass the guard. We do not need this
duke@435	4086	// guard for non-virtual calls; the caller is known to be the native
duke@435	4087	// Object clone().
duke@435	4088	if (is_virtual) {
duke@435	4089	generate_virtual_guard(obj_klass, slow_region);
duke@435	4090	}
duke@435	4091
duke@435	4092	// The object must be cloneable and must not have a finalizer.
duke@435	4093	// Both of these conditions may be checked in a single test.
duke@435	4094	// We could optimize the cloneable test further, but we don't care.
duke@435	4095	generate_access_flags_guard(obj_klass,
duke@435	4096	// Test both conditions:
duke@435	4097	JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
duke@435	4098	// Must be cloneable but not finalizer:
duke@435	4099	JVM_ACC_IS_CLONEABLE,
duke@435	4100	slow_region);
duke@435	4101	}
duke@435	4102
duke@435	4103	if (!stopped()) {
duke@435	4104	// It's an instance, and it passed the slow-path tests.
duke@435	4105	PreserveJVMState pjvms(this);
duke@435	4106	Node* obj_size = NULL;
duke@435	4107	Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
kvn@1268	4108
kvn@1268	4109	copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
kvn@1268	4110
kvn@1268	4111	// Present the results of the slow call.
kvn@1268	4112	result_reg->init_req(_instance_path, control());
kvn@1268	4113	result_val->init_req(_instance_path, alloc_obj);
kvn@1268	4114	result_i_o ->set_req(_instance_path, i_o());
kvn@1268	4115	result_mem ->set_req(_instance_path, reset_memory());
duke@435	4116	}
duke@435	4117
duke@435	4118	// Generate code for the slow case. We make a call to clone().
duke@435	4119	set_control(_gvn.transform(slow_region));
duke@435	4120	if (!stopped()) {
duke@435	4121	PreserveJVMState pjvms(this);
duke@435	4122	CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
duke@435	4123	Node* slow_result = set_results_for_java_call(slow_call);
duke@435	4124	// this->control() comes from set_results_for_java_call
duke@435	4125	result_reg->init_req(_slow_path, control());
duke@435	4126	result_val->init_req(_slow_path, slow_result);
duke@435	4127	result_i_o ->set_req(_slow_path, i_o());
duke@435	4128	result_mem ->set_req(_slow_path, reset_memory());
duke@435	4129	}
duke@435	4130
duke@435	4131	// Return the combined state.
duke@435	4132	set_control( _gvn.transform(result_reg) );
duke@435	4133	set_i_o( _gvn.transform(result_i_o) );
duke@435	4134	set_all_memory( _gvn.transform(result_mem) );
duke@435	4135
kvn@1268	4136	push(_gvn.transform(result_val));
duke@435	4137
duke@435	4138	return true;
duke@435	4139	}
duke@435	4140
duke@435	4141
duke@435	4142	// constants for computing the copy function
duke@435	4143	enum {
duke@435	4144	COPYFUNC_UNALIGNED = 0,
duke@435	4145	COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
duke@435	4146	COPYFUNC_CONJOINT = 0,
duke@435	4147	COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
duke@435	4148	};
duke@435	4149
duke@435	4150	// Note: The condition "disjoint" applies also for overlapping copies
duke@435	4151	// where an descending copy is permitted (i.e., dest_offset <= src_offset).
duke@435	4152	static address
duke@435	4153	select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) {
duke@435	4154	int selector =
duke@435	4155	(aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
duke@435	4156	(disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
duke@435	4157
duke@435	4158	#define RETURN_STUB(xxx_arraycopy) { \
duke@435	4159	name = #xxx_arraycopy; \
duke@435	4160	return StubRoutines::xxx_arraycopy(); }
duke@435	4161
duke@435	4162	switch (t) {
duke@435	4163	case T_BYTE:
duke@435	4164	case T_BOOLEAN:
duke@435	4165	switch (selector) {
duke@435	4166	case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
duke@435	4167	case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
duke@435	4168	case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
duke@435	4169	case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
duke@435	4170	}
duke@435	4171	case T_CHAR:
duke@435	4172	case T_SHORT:
duke@435	4173	switch (selector) {
duke@435	4174	case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
duke@435	4175	case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
duke@435	4176	case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
duke@435	4177	case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
duke@435	4178	}
duke@435	4179	case T_INT:
duke@435	4180	case T_FLOAT:
duke@435	4181	switch (selector) {
duke@435	4182	case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
duke@435	4183	case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
duke@435	4184	case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
duke@435	4185	case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
duke@435	4186	}
duke@435	4187	case T_DOUBLE:
duke@435	4188	case T_LONG:
duke@435	4189	switch (selector) {
duke@435	4190	case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
duke@435	4191	case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
duke@435	4192	case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
duke@435	4193	case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
duke@435	4194	}
duke@435	4195	case T_ARRAY:
duke@435	4196	case T_OBJECT:
duke@435	4197	switch (selector) {
duke@435	4198	case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy);
duke@435	4199	case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy);
duke@435	4200	case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy);
duke@435	4201	case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy);
duke@435	4202	}
duke@435	4203	default:
duke@435	4204	ShouldNotReachHere();
duke@435	4205	return NULL;
duke@435	4206	}
duke@435	4207
duke@435	4208	#undef RETURN_STUB
duke@435	4209	}
duke@435	4210
duke@435	4211	//------------------------------basictype2arraycopy----------------------------
duke@435	4212	address LibraryCallKit::basictype2arraycopy(BasicType t,
duke@435	4213	Node* src_offset,
duke@435	4214	Node* dest_offset,
duke@435	4215	bool disjoint_bases,
duke@435	4216	const char* &name) {
duke@435	4217	const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
duke@435	4218	const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
duke@435	4219
duke@435	4220	bool aligned = false;
duke@435	4221	bool disjoint = disjoint_bases;
duke@435	4222
duke@435	4223	// if the offsets are the same, we can treat the memory regions as
duke@435	4224	// disjoint, because either the memory regions are in different arrays,
duke@435	4225	// or they are identical (which we can treat as disjoint.) We can also
duke@435	4226	// treat a copy with a destination index less that the source index
duke@435	4227	// as disjoint since a low->high copy will work correctly in this case.
duke@435	4228	if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
duke@435	4229	dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
duke@435	4230	// both indices are constants
duke@435	4231	int s_offs = src_offset_inttype->get_con();
duke@435	4232	int d_offs = dest_offset_inttype->get_con();
kvn@464	4233	int element_size = type2aelembytes(t);
duke@435	4234	aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
duke@435	4235	((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
duke@435	4236	if (s_offs >= d_offs) disjoint = true;
duke@435	4237	} else if (src_offset == dest_offset && src_offset != NULL) {
duke@435	4238	// This can occur if the offsets are identical non-constants.
duke@435	4239	disjoint = true;
duke@435	4240	}
duke@435	4241
duke@435	4242	return select_arraycopy_function(t, aligned, disjoint, name);
duke@435	4243	}
duke@435	4244
duke@435	4245
duke@435	4246	//------------------------------inline_arraycopy-----------------------
duke@435	4247	bool LibraryCallKit::inline_arraycopy() {
duke@435	4248	// Restore the stack and pop off the arguments.
duke@435	4249	int nargs = 5; // 2 oops, 3 ints, no size_t or long
duke@435	4250	assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
duke@435	4251
duke@435	4252	Node *src = argument(0);
duke@435	4253	Node *src_offset = argument(1);
duke@435	4254	Node *dest = argument(2);
duke@435	4255	Node *dest_offset = argument(3);
duke@435	4256	Node *length = argument(4);
duke@435	4257
duke@435	4258	// Compile time checks. If any of these checks cannot be verified at compile time,
duke@435	4259	// we do not make a fast path for this call. Instead, we let the call remain as it
duke@435	4260	// is. The checks we choose to mandate at compile time are:
duke@435	4261	//
duke@435	4262	// (1) src and dest are arrays.
duke@435	4263	const Type* src_type = src->Value(&_gvn);
duke@435	4264	const Type* dest_type = dest->Value(&_gvn);
duke@435	4265	const TypeAryPtr* top_src = src_type->isa_aryptr();
duke@435	4266	const TypeAryPtr* top_dest = dest_type->isa_aryptr();
duke@435	4267	if (top_src == NULL || top_src->klass() == NULL ||
duke@435	4268	top_dest == NULL || top_dest->klass() == NULL) {
duke@435	4269	// Conservatively insert a memory barrier on all memory slices.
duke@435	4270	// Do not let writes into the source float below the arraycopy.
duke@435	4271	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	4272
duke@435	4273	// Call StubRoutines::generic_arraycopy stub.
duke@435	4274	generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
kvn@1268	4275	src, src_offset, dest, dest_offset, length);
duke@435	4276
duke@435	4277	// Do not let reads from the destination float above the arraycopy.
duke@435	4278	// Since we cannot type the arrays, we don't know which slices
duke@435	4279	// might be affected. We could restrict this barrier only to those
duke@435	4280	// memory slices which pertain to array elements--but don't bother.
duke@435	4281	if (!InsertMemBarAfterArraycopy)
duke@435	4282	// (If InsertMemBarAfterArraycopy, there is already one in place.)
duke@435	4283	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	4284	return true;
duke@435	4285	}
duke@435	4286
duke@435	4287	// (2) src and dest arrays must have elements of the same BasicType
duke@435	4288	// Figure out the size and type of the elements we will be copying.
duke@435	4289	BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
duke@435	4290	BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
duke@435	4291	if (src_elem == T_ARRAY) src_elem = T_OBJECT;
duke@435	4292	if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
duke@435	4293
duke@435	4294	if (src_elem != dest_elem || dest_elem == T_VOID) {
duke@435	4295	// The component types are not the same or are not recognized. Punt.
duke@435	4296	// (But, avoid the native method wrapper to JVM_ArrayCopy.)
duke@435	4297	generate_slow_arraycopy(TypePtr::BOTTOM,
kvn@1268	4298	src, src_offset, dest, dest_offset, length);
duke@435	4299	return true;
duke@435	4300	}
duke@435	4301
duke@435	4302	//---------------------------------------------------------------------------
duke@435	4303	// We will make a fast path for this call to arraycopy.
duke@435	4304
duke@435	4305	// We have the following tests left to perform:
duke@435	4306	//
duke@435	4307	// (3) src and dest must not be null.
duke@435	4308	// (4) src_offset must not be negative.
duke@435	4309	// (5) dest_offset must not be negative.
duke@435	4310	// (6) length must not be negative.
duke@435	4311	// (7) src_offset + length must not exceed length of src.
duke@435	4312	// (8) dest_offset + length must not exceed length of dest.
duke@435	4313	// (9) each element of an oop array must be assignable
duke@435	4314
duke@435	4315	RegionNode* slow_region = new (C, 1) RegionNode(1);
duke@435	4316	record_for_igvn(slow_region);
duke@435	4317
duke@435	4318	// (3) operands must not be null
duke@435	4319	// We currently perform our null checks with the do_null_check routine.
duke@435	4320	// This means that the null exceptions will be reported in the caller
duke@435	4321	// rather than (correctly) reported inside of the native arraycopy call.
duke@435	4322	// This should be corrected, given time. We do our null check with the
duke@435	4323	// stack pointer restored.
duke@435	4324	_sp += nargs;
duke@435	4325	src = do_null_check(src, T_ARRAY);
duke@435	4326	dest = do_null_check(dest, T_ARRAY);
duke@435	4327	_sp -= nargs;
duke@435	4328
duke@435	4329	// (4) src_offset must not be negative.
duke@435	4330	generate_negative_guard(src_offset, slow_region);
duke@435	4331
duke@435	4332	// (5) dest_offset must not be negative.
duke@435	4333	generate_negative_guard(dest_offset, slow_region);
duke@435	4334
duke@435	4335	// (6) length must not be negative (moved to generate_arraycopy()).
duke@435	4336	// generate_negative_guard(length, slow_region);
duke@435	4337
duke@435	4338	// (7) src_offset + length must not exceed length of src.
duke@435	4339	generate_limit_guard(src_offset, length,
duke@435	4340	load_array_length(src),
duke@435	4341	slow_region);
duke@435	4342
duke@435	4343	// (8) dest_offset + length must not exceed length of dest.
duke@435	4344	generate_limit_guard(dest_offset, length,
duke@435	4345	load_array_length(dest),
duke@435	4346	slow_region);
duke@435	4347
duke@435	4348	// (9) each element of an oop array must be assignable
duke@435	4349	// The generate_arraycopy subroutine checks this.
duke@435	4350
duke@435	4351	// This is where the memory effects are placed:
duke@435	4352	const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
duke@435	4353	generate_arraycopy(adr_type, dest_elem,
duke@435	4354	src, src_offset, dest, dest_offset, length,
kvn@1268	4355	false, false, slow_region);
duke@435	4356
duke@435	4357	return true;
duke@435	4358	}
duke@435	4359
duke@435	4360	//-----------------------------generate_arraycopy----------------------
duke@435	4361	// Generate an optimized call to arraycopy.
duke@435	4362	// Caller must guard against non-arrays.
duke@435	4363	// Caller must determine a common array basic-type for both arrays.
duke@435	4364	// Caller must validate offsets against array bounds.
duke@435	4365	// The slow_region has already collected guard failure paths
duke@435	4366	// (such as out of bounds length or non-conformable array types).
duke@435	4367	// The generated code has this shape, in general:
duke@435	4368	//
duke@435	4369	// if (length == 0) return // via zero_path
duke@435	4370	// slowval = -1
duke@435	4371	// if (types unknown) {
duke@435	4372	// slowval = call generic copy loop
duke@435	4373	// if (slowval == 0) return // via checked_path
duke@435	4374	// } else if (indexes in bounds) {
duke@435	4375	// if ((is object array) && !(array type check)) {
duke@435	4376	// slowval = call checked copy loop
duke@435	4377	// if (slowval == 0) return // via checked_path
duke@435	4378	// } else {
duke@435	4379	// call bulk copy loop
duke@435	4380	// return // via fast_path
duke@435	4381	// }
duke@435	4382	// }
duke@435	4383	// // adjust params for remaining work:
duke@435	4384	// if (slowval != -1) {
duke@435	4385	// n = -1^slowval; src_offset += n; dest_offset += n; length -= n
duke@435	4386	// }
duke@435	4387	// slow_region:
duke@435	4388	// call slow arraycopy(src, src_offset, dest, dest_offset, length)
duke@435	4389	// return // via slow_call_path
duke@435	4390	//
duke@435	4391	// This routine is used from several intrinsics: System.arraycopy,
duke@435	4392	// Object.clone (the array subcase), and Arrays.copyOf[Range].
duke@435	4393	//
duke@435	4394	void
duke@435	4395	LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
duke@435	4396	BasicType basic_elem_type,
duke@435	4397	Node* src, Node* src_offset,
duke@435	4398	Node* dest, Node* dest_offset,
duke@435	4399	Node* copy_length,
duke@435	4400	bool disjoint_bases,
duke@435	4401	bool length_never_negative,
duke@435	4402	RegionNode* slow_region) {
duke@435	4403
duke@435	4404	if (slow_region == NULL) {
duke@435	4405	slow_region = new(C,1) RegionNode(1);
duke@435	4406	record_for_igvn(slow_region);
duke@435	4407	}
duke@435	4408
duke@435	4409	Node* original_dest = dest;
duke@435	4410	AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
duke@435	4411	bool must_clear_dest = false;
duke@435	4412
duke@435	4413	// See if this is the initialization of a newly-allocated array.
duke@435	4414	// If so, we will take responsibility here for initializing it to zero.
duke@435	4415	// (Note: Because tightly_coupled_allocation performs checks on the
duke@435	4416	// out-edges of the dest, we need to avoid making derived pointers
duke@435	4417	// from it until we have checked its uses.)
duke@435	4418	if (ReduceBulkZeroing
duke@435	4419	&& !ZeroTLAB // pointless if already zeroed
duke@435	4420	&& basic_elem_type != T_CONFLICT // avoid corner case
duke@435	4421	&& !_gvn.eqv_uncast(src, dest)
duke@435	4422	&& ((alloc = tightly_coupled_allocation(dest, slow_region))
duke@435	4423	!= NULL)
kvn@469	4424	&& _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
duke@435	4425	&& alloc->maybe_set_complete(&_gvn)) {
duke@435	4426	// "You break it, you buy it."
duke@435	4427	InitializeNode* init = alloc->initialization();
duke@435	4428	assert(init->is_complete(), "we just did this");
kvn@1268	4429	assert(dest->is_CheckCastPP(), "sanity");
duke@435	4430	assert(dest->in(0)->in(0) == init, "dest pinned");
kvn@1268	4431
kvn@1268	4432	// Cast to Object for arraycopy.
kvn@1268	4433	// We can't use the original CheckCastPP since it should be moved
kvn@1268	4434	// after the arraycopy to prevent stores flowing above it.
kvn@1268	4435	Node* new_obj = new(C, 2) CheckCastPPNode(dest->in(0), dest->in(1),
kvn@1268	4436	TypeInstPtr::NOTNULL);
kvn@1268	4437	dest = _gvn.transform(new_obj);
kvn@1268	4438	// Substitute in the locally valid dest_oop.
kvn@1268	4439	replace_in_map(original_dest, dest);
duke@435	4440	adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
duke@435	4441	// From this point on, every exit path is responsible for
duke@435	4442	// initializing any non-copied parts of the object to zero.
duke@435	4443	must_clear_dest = true;
duke@435	4444	} else {
duke@435	4445	// No zeroing elimination here.
duke@435	4446	alloc = NULL;
duke@435	4447	//original_dest = dest;
duke@435	4448	//must_clear_dest = false;
duke@435	4449	}
duke@435	4450
duke@435	4451	// Results are placed here:
duke@435	4452	enum { fast_path = 1, // normal void-returning assembly stub
duke@435	4453	checked_path = 2, // special assembly stub with cleanup
duke@435	4454	slow_call_path = 3, // something went wrong; call the VM
duke@435	4455	zero_path = 4, // bypass when length of copy is zero
duke@435	4456	bcopy_path = 5, // copy primitive array by 64-bit blocks
duke@435	4457	PATH_LIMIT = 6
duke@435	4458	};
duke@435	4459	RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
duke@435	4460	PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
duke@435	4461	PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
duke@435	4462	record_for_igvn(result_region);
duke@435	4463	_gvn.set_type_bottom(result_i_o);
duke@435	4464	_gvn.set_type_bottom(result_memory);
duke@435	4465	assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
duke@435	4466
duke@435	4467	// The slow_control path:
duke@435	4468	Node* slow_control;
duke@435	4469	Node* slow_i_o = i_o();
duke@435	4470	Node* slow_mem = memory(adr_type);
duke@435	4471	debug_only(slow_control = (Node*) badAddress);
duke@435	4472
duke@435	4473	// Checked control path:
duke@435	4474	Node* checked_control = top();
duke@435	4475	Node* checked_mem = NULL;
duke@435	4476	Node* checked_i_o = NULL;
duke@435	4477	Node* checked_value = NULL;
duke@435	4478
duke@435	4479	if (basic_elem_type == T_CONFLICT) {
duke@435	4480	assert(!must_clear_dest, "");
duke@435	4481	Node* cv = generate_generic_arraycopy(adr_type,
duke@435	4482	src, src_offset, dest, dest_offset,
kvn@1268	4483	copy_length);
duke@435	4484	if (cv == NULL) cv = intcon(-1); // failure (no stub available)
duke@435	4485	checked_control = control();
duke@435	4486	checked_i_o = i_o();
duke@435	4487	checked_mem = memory(adr_type);
duke@435	4488	checked_value = cv;
duke@435	4489	set_control(top()); // no fast path
duke@435	4490	}
duke@435	4491
duke@435	4492	Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
duke@435	4493	if (not_pos != NULL) {
duke@435	4494	PreserveJVMState pjvms(this);
duke@435	4495	set_control(not_pos);
duke@435	4496
duke@435	4497	// (6) length must not be negative.
duke@435	4498	if (!length_never_negative) {
duke@435	4499	generate_negative_guard(copy_length, slow_region);
duke@435	4500	}
duke@435	4501
kvn@1271	4502	// copy_length is 0.
duke@435	4503	if (!stopped() && must_clear_dest) {
duke@435	4504	Node* dest_length = alloc->in(AllocateNode::ALength);
duke@435	4505	if (_gvn.eqv_uncast(copy_length, dest_length)
duke@435	4506	|| _gvn.find_int_con(dest_length, 1) <= 0) {
kvn@1271	4507	// There is no zeroing to do. No need for a secondary raw memory barrier.
duke@435	4508	} else {
duke@435	4509	// Clear the whole thing since there are no source elements to copy.
duke@435	4510	generate_clear_array(adr_type, dest, basic_elem_type,
duke@435	4511	intcon(0), NULL,
duke@435	4512	alloc->in(AllocateNode::AllocSize));
kvn@1271	4513	// Use a secondary InitializeNode as raw memory barrier.
kvn@1271	4514	// Currently it is needed only on this path since other
kvn@1271	4515	// paths have stub or runtime calls as raw memory barriers.
kvn@1271	4516	InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
kvn@1271	4517	Compile::AliasIdxRaw,
kvn@1271	4518	top())->as_Initialize();
kvn@1271	4519	init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
duke@435	4520	}
duke@435	4521	}
duke@435	4522
duke@435	4523	// Present the results of the fast call.
duke@435	4524	result_region->init_req(zero_path, control());
duke@435	4525	result_i_o ->init_req(zero_path, i_o());
duke@435	4526	result_memory->init_req(zero_path, memory(adr_type));
duke@435	4527	}
duke@435	4528
duke@435	4529	if (!stopped() && must_clear_dest) {
duke@435	4530	// We have to initialize the *uncopied* part of the array to zero.
duke@435	4531	// The copy destination is the slice dest[off..off+len]. The other slices
duke@435	4532	// are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
duke@435	4533	Node* dest_size = alloc->in(AllocateNode::AllocSize);
duke@435	4534	Node* dest_length = alloc->in(AllocateNode::ALength);
duke@435	4535	Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
duke@435	4536	copy_length) );
duke@435	4537
duke@435	4538	// If there is a head section that needs zeroing, do it now.
duke@435	4539	if (find_int_con(dest_offset, -1) != 0) {
duke@435	4540	generate_clear_array(adr_type, dest, basic_elem_type,
duke@435	4541	intcon(0), dest_offset,
duke@435	4542	NULL);
duke@435	4543	}
duke@435	4544
duke@435	4545	// Next, perform a dynamic check on the tail length.
duke@435	4546	// It is often zero, and we can win big if we prove this.
duke@435	4547	// There are two wins: Avoid generating the ClearArray
duke@435	4548	// with its attendant messy index arithmetic, and upgrade
duke@435	4549	// the copy to a more hardware-friendly word size of 64 bits.
duke@435	4550	Node* tail_ctl = NULL;
duke@435	4551	if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
duke@435	4552	Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
duke@435	4553	Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
duke@435	4554	tail_ctl = generate_slow_guard(bol_lt, NULL);
duke@435	4555	assert(tail_ctl != NULL || !stopped(), "must be an outcome");
duke@435	4556	}
duke@435	4557
duke@435	4558	// At this point, let's assume there is no tail.
duke@435	4559	if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
duke@435	4560	// There is no tail. Try an upgrade to a 64-bit copy.
duke@435	4561	bool didit = false;
duke@435	4562	{ PreserveJVMState pjvms(this);
duke@435	4563	didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
duke@435	4564	src, src_offset, dest, dest_offset,
duke@435	4565	dest_size);
duke@435	4566	if (didit) {
duke@435	4567	// Present the results of the block-copying fast call.
duke@435	4568	result_region->init_req(bcopy_path, control());
duke@435	4569	result_i_o ->init_req(bcopy_path, i_o());
duke@435	4570	result_memory->init_req(bcopy_path, memory(adr_type));
duke@435	4571	}
duke@435	4572	}
duke@435	4573	if (didit)
duke@435	4574	set_control(top()); // no regular fast path
duke@435	4575	}
duke@435	4576
duke@435	4577	// Clear the tail, if any.
duke@435	4578	if (tail_ctl != NULL) {
duke@435	4579	Node* notail_ctl = stopped() ? NULL : control();
duke@435	4580	set_control(tail_ctl);
duke@435	4581	if (notail_ctl == NULL) {
duke@435	4582	generate_clear_array(adr_type, dest, basic_elem_type,
duke@435	4583	dest_tail, NULL,
duke@435	4584	dest_size);
duke@435	4585	} else {
duke@435	4586	// Make a local merge.
duke@435	4587	Node* done_ctl = new(C,3) RegionNode(3);
duke@435	4588	Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
duke@435	4589	done_ctl->init_req(1, notail_ctl);
duke@435	4590	done_mem->init_req(1, memory(adr_type));
duke@435	4591	generate_clear_array(adr_type, dest, basic_elem_type,
duke@435	4592	dest_tail, NULL,
duke@435	4593	dest_size);
duke@435	4594	done_ctl->init_req(2, control());
duke@435	4595	done_mem->init_req(2, memory(adr_type));
duke@435	4596	set_control( _gvn.transform(done_ctl) );
duke@435	4597	set_memory( _gvn.transform(done_mem), adr_type );
duke@435	4598	}
duke@435	4599	}
duke@435	4600	}
duke@435	4601
duke@435	4602	BasicType copy_type = basic_elem_type;
duke@435	4603	assert(basic_elem_type != T_ARRAY, "caller must fix this");
duke@435	4604	if (!stopped() && copy_type == T_OBJECT) {
duke@435	4605	// If src and dest have compatible element types, we can copy bits.
duke@435	4606	// Types S[] and D[] are compatible if D is a supertype of S.
duke@435	4607	//
duke@435	4608	// If they are not, we will use checked_oop_disjoint_arraycopy,
duke@435	4609	// which performs a fast optimistic per-oop check, and backs off
duke@435	4610	// further to JVM_ArrayCopy on the first per-oop check that fails.
duke@435	4611	// (Actually, we don't move raw bits only; the GC requires card marks.)
duke@435	4612
duke@435	4613	// Get the klassOop for both src and dest
duke@435	4614	Node* src_klass = load_object_klass(src);
duke@435	4615	Node* dest_klass = load_object_klass(dest);
duke@435	4616
duke@435	4617	// Generate the subtype check.
duke@435	4618	// This might fold up statically, or then again it might not.
duke@435	4619	//
duke@435	4620	// Non-static example: Copying List<String>.elements to a new String[].
duke@435	4621	// The backing store for a List<String> is always an Object[],
duke@435	4622	// but its elements are always type String, if the generic types
duke@435	4623	// are correct at the source level.
duke@435	4624	//
duke@435	4625	// Test S[] against D[], not S against D, because (probably)
duke@435	4626	// the secondary supertype cache is less busy for S[] than S.
duke@435	4627	// This usually only matters when D is an interface.
duke@435	4628	Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
duke@435	4629	// Plug failing path into checked_oop_disjoint_arraycopy
duke@435	4630	if (not_subtype_ctrl != top()) {
duke@435	4631	PreserveJVMState pjvms(this);
duke@435	4632	set_control(not_subtype_ctrl);
duke@435	4633	// (At this point we can assume disjoint_bases, since types differ.)
duke@435	4634	int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
duke@435	4635	Node* p1 = basic_plus_adr(dest_klass, ek_offset);
kvn@599	4636	Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
duke@435	4637	Node* dest_elem_klass = _gvn.transform(n1);
duke@435	4638	Node* cv = generate_checkcast_arraycopy(adr_type,
duke@435	4639	dest_elem_klass,
duke@435	4640	src, src_offset, dest, dest_offset,
kvn@1268	4641	copy_length);
duke@435	4642	if (cv == NULL) cv = intcon(-1); // failure (no stub available)
duke@435	4643	checked_control = control();
duke@435	4644	checked_i_o = i_o();
duke@435	4645	checked_mem = memory(adr_type);
duke@435	4646	checked_value = cv;
duke@435	4647	}
duke@435	4648	// At this point we know we do not need type checks on oop stores.
duke@435	4649
duke@435	4650	// Let's see if we need card marks:
duke@435	4651	if (alloc != NULL && use_ReduceInitialCardMarks()) {
duke@435	4652	// If we do not need card marks, copy using the jint or jlong stub.
coleenp@548	4653	copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
kvn@464	4654	assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
duke@435	4655	"sizes agree");
duke@435	4656	}
duke@435	4657	}
duke@435	4658
duke@435	4659	if (!stopped()) {
duke@435	4660	// Generate the fast path, if possible.
duke@435	4661	PreserveJVMState pjvms(this);
duke@435	4662	generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
duke@435	4663	src, src_offset, dest, dest_offset,
duke@435	4664	ConvI2X(copy_length));
duke@435	4665
duke@435	4666	// Present the results of the fast call.
duke@435	4667	result_region->init_req(fast_path, control());
duke@435	4668	result_i_o ->init_req(fast_path, i_o());
duke@435	4669	result_memory->init_req(fast_path, memory(adr_type));
duke@435	4670	}
duke@435	4671
duke@435	4672	// Here are all the slow paths up to this point, in one bundle:
duke@435	4673	slow_control = top();
duke@435	4674	if (slow_region != NULL)
duke@435	4675	slow_control = _gvn.transform(slow_region);
duke@435	4676	debug_only(slow_region = (RegionNode*)badAddress);
duke@435	4677
duke@435	4678	set_control(checked_control);
duke@435	4679	if (!stopped()) {
duke@435	4680	// Clean up after the checked call.
duke@435	4681	// The returned value is either 0 or -1^K,
duke@435	4682	// where K = number of partially transferred array elements.
duke@435	4683	Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
duke@435	4684	Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
duke@435	4685	IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
duke@435	4686
duke@435	4687	// If it is 0, we are done, so transfer to the end.
duke@435	4688	Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
duke@435	4689	result_region->init_req(checked_path, checks_done);
duke@435	4690	result_i_o ->init_req(checked_path, checked_i_o);
duke@435	4691	result_memory->init_req(checked_path, checked_mem);
duke@435	4692
duke@435	4693	// If it is not zero, merge into the slow call.
duke@435	4694	set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
duke@435	4695	RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
duke@435	4696	PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
duke@435	4697	PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
duke@435	4698	record_for_igvn(slow_reg2);
duke@435	4699	slow_reg2 ->init_req(1, slow_control);
duke@435	4700	slow_i_o2 ->init_req(1, slow_i_o);
duke@435	4701	slow_mem2 ->init_req(1, slow_mem);
duke@435	4702	slow_reg2 ->init_req(2, control());
kvn@1268	4703	slow_i_o2 ->init_req(2, checked_i_o);
kvn@1268	4704	slow_mem2 ->init_req(2, checked_mem);
duke@435	4705
duke@435	4706	slow_control = _gvn.transform(slow_reg2);
duke@435	4707	slow_i_o = _gvn.transform(slow_i_o2);
duke@435	4708	slow_mem = _gvn.transform(slow_mem2);
duke@435	4709
duke@435	4710	if (alloc != NULL) {
duke@435	4711	// We'll restart from the very beginning, after zeroing the whole thing.
duke@435	4712	// This can cause double writes, but that's OK since dest is brand new.
duke@435	4713	// So we ignore the low 31 bits of the value returned from the stub.
duke@435	4714	} else {
duke@435	4715	// We must continue the copy exactly where it failed, or else
duke@435	4716	// another thread might see the wrong number of writes to dest.
duke@435	4717	Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
duke@435	4718	Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
duke@435	4719	slow_offset->init_req(1, intcon(0));
duke@435	4720	slow_offset->init_req(2, checked_offset);
duke@435	4721	slow_offset = _gvn.transform(slow_offset);
duke@435	4722
duke@435	4723	// Adjust the arguments by the conditionally incoming offset.
duke@435	4724	Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
duke@435	4725	Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
duke@435	4726	Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
duke@435	4727
duke@435	4728	// Tweak the node variables to adjust the code produced below:
duke@435	4729	src_offset = src_off_plus;
duke@435	4730	dest_offset = dest_off_plus;
duke@435	4731	copy_length = length_minus;
duke@435	4732	}
duke@435	4733	}
duke@435	4734
duke@435	4735	set_control(slow_control);
duke@435	4736	if (!stopped()) {
duke@435	4737	// Generate the slow path, if needed.
duke@435	4738	PreserveJVMState pjvms(this); // replace_in_map may trash the map
duke@435	4739
duke@435	4740	set_memory(slow_mem, adr_type);
duke@435	4741	set_i_o(slow_i_o);
duke@435	4742
duke@435	4743	if (must_clear_dest) {
duke@435	4744	generate_clear_array(adr_type, dest, basic_elem_type,
duke@435	4745	intcon(0), NULL,
duke@435	4746	alloc->in(AllocateNode::AllocSize));
duke@435	4747	}
duke@435	4748
duke@435	4749	generate_slow_arraycopy(adr_type,
duke@435	4750	src, src_offset, dest, dest_offset,
kvn@1268	4751	copy_length);
duke@435	4752
duke@435	4753	result_region->init_req(slow_call_path, control());
duke@435	4754	result_i_o ->init_req(slow_call_path, i_o());
duke@435	4755	result_memory->init_req(slow_call_path, memory(adr_type));
duke@435	4756	}
duke@435	4757
duke@435	4758	// Remove unused edges.
duke@435	4759	for (uint i = 1; i < result_region->req(); i++) {
duke@435	4760	if (result_region->in(i) == NULL)
duke@435	4761	result_region->init_req(i, top());
duke@435	4762	}
duke@435	4763
duke@435	4764	// Finished; return the combined state.
duke@435	4765	set_control( _gvn.transform(result_region) );
duke@435	4766	set_i_o( _gvn.transform(result_i_o) );
duke@435	4767	set_memory( _gvn.transform(result_memory), adr_type );
duke@435	4768
duke@435	4769	if (dest != original_dest) {
duke@435	4770	// Pin the "finished" array node after the arraycopy/zeroing operations.
duke@435	4771	_gvn.hash_delete(original_dest);
duke@435	4772	original_dest->set_req(0, control());
kvn@1268	4773	// Replace raw memory edge with new CheckCastPP to have a live oop
kvn@1268	4774	// at safepoints instead of raw value.
kvn@1268	4775	assert(dest->is_CheckCastPP() && dest->in(1) == original_dest->in(1), "sanity");
kvn@1268	4776	original_dest->set_req(1, dest); // cast to the original type
duke@435	4777	_gvn.hash_find_insert(original_dest); // put back into GVN table
kvn@1268	4778	// Restore in the locally valid dest_oop.
kvn@1268	4779	replace_in_map(dest, original_dest);
duke@435	4780	}
duke@435	4781	// The memory edges above are precise in order to model effects around
twisti@1040	4782	// array copies accurately to allow value numbering of field loads around
duke@435	4783	// arraycopy. Such field loads, both before and after, are common in Java
duke@435	4784	// collections and similar classes involving header/array data structures.
duke@435	4785	//
duke@435	4786	// But with low number of register or when some registers are used or killed
duke@435	4787	// by arraycopy calls it causes registers spilling on stack. See 6544710.
duke@435	4788	// The next memory barrier is added to avoid it. If the arraycopy can be
duke@435	4789	// optimized away (which it can, sometimes) then we can manually remove
duke@435	4790	// the membar also.
duke@435	4791	if (InsertMemBarAfterArraycopy)
duke@435	4792	insert_mem_bar(Op_MemBarCPUOrder);
duke@435	4793	}
duke@435	4794
duke@435	4795
duke@435	4796	// Helper function which determines if an arraycopy immediately follows
duke@435	4797	// an allocation, with no intervening tests or other escapes for the object.
duke@435	4798	AllocateArrayNode*
duke@435	4799	LibraryCallKit::tightly_coupled_allocation(Node* ptr,
duke@435	4800	RegionNode* slow_region) {
duke@435	4801	if (stopped()) return NULL; // no fast path
duke@435	4802	if (C->AliasLevel() == 0) return NULL; // no MergeMems around
duke@435	4803
duke@435	4804	AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
duke@435	4805	if (alloc == NULL) return NULL;
duke@435	4806
duke@435	4807	Node* rawmem = memory(Compile::AliasIdxRaw);
duke@435	4808	// Is the allocation's memory state untouched?
duke@435	4809	if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
duke@435	4810	// Bail out if there have been raw-memory effects since the allocation.
duke@435	4811	// (Example: There might have been a call or safepoint.)
duke@435	4812	return NULL;
duke@435	4813	}
duke@435	4814	rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
duke@435	4815	if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
duke@435	4816	return NULL;
duke@435	4817	}
duke@435	4818
duke@435	4819	// There must be no unexpected observers of this allocation.
duke@435	4820	for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
duke@435	4821	Node* obs = ptr->fast_out(i);
duke@435	4822	if (obs != this->map()) {
duke@435	4823	return NULL;
duke@435	4824	}
duke@435	4825	}
duke@435	4826
duke@435	4827	// This arraycopy must unconditionally follow the allocation of the ptr.
duke@435	4828	Node* alloc_ctl = ptr->in(0);
duke@435	4829	assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
duke@435	4830
duke@435	4831	Node* ctl = control();
duke@435	4832	while (ctl != alloc_ctl) {
duke@435	4833	// There may be guards which feed into the slow_region.
duke@435	4834	// Any other control flow means that we might not get a chance
duke@435	4835	// to finish initializing the allocated object.
duke@435	4836	if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
duke@435	4837	IfNode* iff = ctl->in(0)->as_If();
duke@435	4838	Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
duke@435	4839	assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
duke@435	4840	if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
duke@435	4841	ctl = iff->in(0); // This test feeds the known slow_region.
duke@435	4842	continue;
duke@435	4843	}
duke@435	4844	// One more try: Various low-level checks bottom out in
duke@435	4845	// uncommon traps. If the debug-info of the trap omits
duke@435	4846	// any reference to the allocation, as we've already
duke@435	4847	// observed, then there can be no objection to the trap.
duke@435	4848	bool found_trap = false;
duke@435	4849	for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
duke@435	4850	Node* obs = not_ctl->fast_out(j);
duke@435	4851	if (obs->in(0) == not_ctl && obs->is_Call() &&
duke@435	4852	(obs->as_Call()->entry_point() ==
duke@435	4853	SharedRuntime::uncommon_trap_blob()->instructions_begin())) {
duke@435	4854	found_trap = true; break;
duke@435	4855	}
duke@435	4856	}
duke@435	4857	if (found_trap) {
duke@435	4858	ctl = iff->in(0); // This test feeds a harmless uncommon trap.
duke@435	4859	continue;
duke@435	4860	}
duke@435	4861	}
duke@435	4862	return NULL;
duke@435	4863	}
duke@435	4864
duke@435	4865	// If we get this far, we have an allocation which immediately
duke@435	4866	// precedes the arraycopy, and we can take over zeroing the new object.
duke@435	4867	// The arraycopy will finish the initialization, and provide
duke@435	4868	// a new control state to which we will anchor the destination pointer.
duke@435	4869
duke@435	4870	return alloc;
duke@435	4871	}
duke@435	4872
duke@435	4873	// Helper for initialization of arrays, creating a ClearArray.
duke@435	4874	// It writes zero bits in [start..end), within the body of an array object.
duke@435	4875	// The memory effects are all chained onto the 'adr_type' alias category.
duke@435	4876	//
duke@435	4877	// Since the object is otherwise uninitialized, we are free
duke@435	4878	// to put a little "slop" around the edges of the cleared area,
duke@435	4879	// as long as it does not go back into the array's header,
duke@435	4880	// or beyond the array end within the heap.
duke@435	4881	//
duke@435	4882	// The lower edge can be rounded down to the nearest jint and the
duke@435	4883	// upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
duke@435	4884	//
duke@435	4885	// Arguments:
duke@435	4886	// adr_type memory slice where writes are generated
duke@435	4887	// dest oop of the destination array
duke@435	4888	// basic_elem_type element type of the destination
duke@435	4889	// slice_idx array index of first element to store
duke@435	4890	// slice_len number of elements to store (or NULL)
duke@435	4891	// dest_size total size in bytes of the array object
duke@435	4892	//
duke@435	4893	// Exactly one of slice_len or dest_size must be non-NULL.
duke@435	4894	// If dest_size is non-NULL, zeroing extends to the end of the object.
duke@435	4895	// If slice_len is non-NULL, the slice_idx value must be a constant.
duke@435	4896	void
duke@435	4897	LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
duke@435	4898	Node* dest,
duke@435	4899	BasicType basic_elem_type,
duke@435	4900	Node* slice_idx,
duke@435	4901	Node* slice_len,
duke@435	4902	Node* dest_size) {
duke@435	4903	// one or the other but not both of slice_len and dest_size:
duke@435	4904	assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
duke@435	4905	if (slice_len == NULL) slice_len = top();
duke@435	4906	if (dest_size == NULL) dest_size = top();
duke@435	4907
duke@435	4908	// operate on this memory slice:
duke@435	4909	Node* mem = memory(adr_type); // memory slice to operate on
duke@435	4910
duke@435	4911	// scaling and rounding of indexes:
kvn@464	4912	int scale = exact_log2(type2aelembytes(basic_elem_type));
duke@435	4913	int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
duke@435	4914	int clear_low = (-1 << scale) & (BytesPerInt - 1);
duke@435	4915	int bump_bit = (-1 << scale) & BytesPerInt;
duke@435	4916
duke@435	4917	// determine constant starts and ends
duke@435	4918	const intptr_t BIG_NEG = -128;
duke@435	4919	assert(BIG_NEG + 2*abase < 0, "neg enough");
duke@435	4920	intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
duke@435	4921	intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
duke@435	4922	if (slice_len_con == 0) {
duke@435	4923	return; // nothing to do here
duke@435	4924	}
duke@435	4925	intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
duke@435	4926	intptr_t end_con = find_intptr_t_con(dest_size, -1);
duke@435	4927	if (slice_idx_con >= 0 && slice_len_con >= 0) {
duke@435	4928	assert(end_con < 0, "not two cons");
duke@435	4929	end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
duke@435	4930	BytesPerLong);
duke@435	4931	}
duke@435	4932
duke@435	4933	if (start_con >= 0 && end_con >= 0) {
duke@435	4934	// Constant start and end. Simple.
duke@435	4935	mem = ClearArrayNode::clear_memory(control(), mem, dest,
duke@435	4936	start_con, end_con, &_gvn);
duke@435	4937	} else if (start_con >= 0 && dest_size != top()) {
duke@435	4938	// Constant start, pre-rounded end after the tail of the array.
duke@435	4939	Node* end = dest_size;
duke@435	4940	mem = ClearArrayNode::clear_memory(control(), mem, dest,
duke@435	4941	start_con, end, &_gvn);
duke@435	4942	} else if (start_con >= 0 && slice_len != top()) {
duke@435	4943	// Constant start, non-constant end. End needs rounding up.
duke@435	4944	// End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
duke@435	4945	intptr_t end_base = abase + (slice_idx_con << scale);
duke@435	4946	int end_round = (-1 << scale) & (BytesPerLong - 1);
duke@435	4947	Node* end = ConvI2X(slice_len);
duke@435	4948	if (scale != 0)
duke@435	4949	end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
duke@435	4950	end_base += end_round;
duke@435	4951	end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
duke@435	4952	end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
duke@435	4953	mem = ClearArrayNode::clear_memory(control(), mem, dest,
duke@435	4954	start_con, end, &_gvn);
duke@435	4955	} else if (start_con < 0 && dest_size != top()) {
duke@435	4956	// Non-constant start, pre-rounded end after the tail of the array.
duke@435	4957	// This is almost certainly a "round-to-end" operation.
duke@435	4958	Node* start = slice_idx;
duke@435	4959	start = ConvI2X(start);
duke@435	4960	if (scale != 0)
duke@435	4961	start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
duke@435	4962	start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
duke@435	4963	if ((bump_bit | clear_low) != 0) {
duke@435	4964	int to_clear = (bump_bit | clear_low);
duke@435	4965	// Align up mod 8, then store a jint zero unconditionally
duke@435	4966	// just before the mod-8 boundary.
coleenp@548	4967	if (((abase + bump_bit) & ~to_clear) - bump_bit
coleenp@548	4968	< arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
coleenp@548	4969	bump_bit = 0;
coleenp@548	4970	assert((abase & to_clear) == 0, "array base must be long-aligned");
coleenp@548	4971	} else {
coleenp@548	4972	// Bump 'start' up to (or past) the next jint boundary:
coleenp@548	4973	start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
coleenp@548	4974	assert((abase & clear_low) == 0, "array base must be int-aligned");
coleenp@548	4975	}
duke@435	4976	// Round bumped 'start' down to jlong boundary in body of array.
duke@435	4977	start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
coleenp@548	4978	if (bump_bit != 0) {
coleenp@548	4979	// Store a zero to the immediately preceding jint:
coleenp@548	4980	Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
coleenp@548	4981	Node* p1 = basic_plus_adr(dest, x1);
coleenp@548	4982	mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
coleenp@548	4983	mem = _gvn.transform(mem);
coleenp@548	4984	}
duke@435	4985	}
duke@435	4986	Node* end = dest_size; // pre-rounded
duke@435	4987	mem = ClearArrayNode::clear_memory(control(), mem, dest,
duke@435	4988	start, end, &_gvn);
duke@435	4989	} else {
duke@435	4990	// Non-constant start, unrounded non-constant end.
duke@435	4991	// (Nobody zeroes a random midsection of an array using this routine.)
duke@435	4992	ShouldNotReachHere(); // fix caller
duke@435	4993	}
duke@435	4994
duke@435	4995	// Done.
duke@435	4996	set_memory(mem, adr_type);
duke@435	4997	}
duke@435	4998
duke@435	4999
duke@435	5000	bool
duke@435	5001	LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
duke@435	5002	BasicType basic_elem_type,
duke@435	5003	AllocateNode* alloc,
duke@435	5004	Node* src, Node* src_offset,
duke@435	5005	Node* dest, Node* dest_offset,
duke@435	5006	Node* dest_size) {
duke@435	5007	// See if there is an advantage from block transfer.
kvn@464	5008	int scale = exact_log2(type2aelembytes(basic_elem_type));
duke@435	5009	if (scale >= LogBytesPerLong)
duke@435	5010	return false; // it is already a block transfer
duke@435	5011
duke@435	5012	// Look at the alignment of the starting offsets.
duke@435	5013	int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
duke@435	5014	const intptr_t BIG_NEG = -128;
duke@435	5015	assert(BIG_NEG + 2*abase < 0, "neg enough");
duke@435	5016
duke@435	5017	intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
duke@435	5018	intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
duke@435	5019	if (src_off < 0 || dest_off < 0)
duke@435	5020	// At present, we can only understand constants.
duke@435	5021	return false;
duke@435	5022
duke@435	5023	if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
duke@435	5024	// Non-aligned; too bad.
duke@435	5025	// One more chance: Pick off an initial 32-bit word.
duke@435	5026	// This is a common case, since abase can be odd mod 8.
duke@435	5027	if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
duke@435	5028	((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
duke@435	5029	Node* sptr = basic_plus_adr(src, src_off);
duke@435	5030	Node* dptr = basic_plus_adr(dest, dest_off);
duke@435	5031	Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
duke@435	5032	store_to_memory(control(), dptr, sval, T_INT, adr_type);
duke@435	5033	src_off += BytesPerInt;
duke@435	5034	dest_off += BytesPerInt;
duke@435	5035	} else {
duke@435	5036	return false;
duke@435	5037	}
duke@435	5038	}
duke@435	5039	assert(src_off % BytesPerLong == 0, "");
duke@435	5040	assert(dest_off % BytesPerLong == 0, "");
duke@435	5041
duke@435	5042	// Do this copy by giant steps.
duke@435	5043	Node* sptr = basic_plus_adr(src, src_off);
duke@435	5044	Node* dptr = basic_plus_adr(dest, dest_off);
duke@435	5045	Node* countx = dest_size;
duke@435	5046	countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
duke@435	5047	countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
duke@435	5048
duke@435	5049	bool disjoint_bases = true; // since alloc != NULL
duke@435	5050	generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
duke@435	5051	sptr, NULL, dptr, NULL, countx);
duke@435	5052
duke@435	5053	return true;
duke@435	5054	}
duke@435	5055
duke@435	5056
duke@435	5057	// Helper function; generates code for the slow case.
duke@435	5058	// We make a call to a runtime method which emulates the native method,
duke@435	5059	// but without the native wrapper overhead.
duke@435	5060	void
duke@435	5061	LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
duke@435	5062	Node* src, Node* src_offset,
duke@435	5063	Node* dest, Node* dest_offset,
kvn@1268	5064	Node* copy_length) {
duke@435	5065	Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
duke@435	5066	OptoRuntime::slow_arraycopy_Type(),
duke@435	5067	OptoRuntime::slow_arraycopy_Java(),
duke@435	5068	"slow_arraycopy", adr_type,
duke@435	5069	src, src_offset, dest, dest_offset,
duke@435	5070	copy_length);
duke@435	5071
duke@435	5072	// Handle exceptions thrown by this fellow:
duke@435	5073	make_slow_call_ex(call, env()->Throwable_klass(), false);
duke@435	5074	}
duke@435	5075
duke@435	5076	// Helper function; generates code for cases requiring runtime checks.
duke@435	5077	Node*
duke@435	5078	LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
duke@435	5079	Node* dest_elem_klass,
duke@435	5080	Node* src, Node* src_offset,
duke@435	5081	Node* dest, Node* dest_offset,
kvn@1268	5082	Node* copy_length) {
duke@435	5083	if (stopped()) return NULL;
duke@435	5084
duke@435	5085	address copyfunc_addr = StubRoutines::checkcast_arraycopy();
duke@435	5086	if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
duke@435	5087	return NULL;
duke@435	5088	}
duke@435	5089
duke@435	5090	// Pick out the parameters required to perform a store-check
duke@435	5091	// for the target array. This is an optimistic check. It will
duke@435	5092	// look in each non-null element's class, at the desired klass's
duke@435	5093	// super_check_offset, for the desired klass.
duke@435	5094	int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
duke@435	5095	Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
duke@435	5096	Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM);
duke@435	5097	Node* check_offset = _gvn.transform(n3);
duke@435	5098	Node* check_value = dest_elem_klass;
duke@435	5099
duke@435	5100	Node* src_start = array_element_address(src, src_offset, T_OBJECT);
duke@435	5101	Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
duke@435	5102
duke@435	5103	// (We know the arrays are never conjoint, because their types differ.)
duke@435	5104	Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
duke@435	5105	OptoRuntime::checkcast_arraycopy_Type(),
duke@435	5106	copyfunc_addr, "checkcast_arraycopy", adr_type,
duke@435	5107	// five arguments, of which two are
duke@435	5108	// intptr_t (jlong in LP64)
duke@435	5109	src_start, dest_start,
duke@435	5110	copy_length XTOP,
duke@435	5111	check_offset XTOP,
duke@435	5112	check_value);
duke@435	5113
duke@435	5114	return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
duke@435	5115	}
duke@435	5116
duke@435	5117
duke@435	5118	// Helper function; generates code for cases requiring runtime checks.
duke@435	5119	Node*
duke@435	5120	LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
duke@435	5121	Node* src, Node* src_offset,
duke@435	5122	Node* dest, Node* dest_offset,
kvn@1268	5123	Node* copy_length) {
duke@435	5124	if (stopped()) return NULL;
duke@435	5125
duke@435	5126	address copyfunc_addr = StubRoutines::generic_arraycopy();
duke@435	5127	if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
duke@435	5128	return NULL;
duke@435	5129	}
duke@435	5130
duke@435	5131	Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
duke@435	5132	OptoRuntime::generic_arraycopy_Type(),
duke@435	5133	copyfunc_addr, "generic_arraycopy", adr_type,
duke@435	5134	src, src_offset, dest, dest_offset, copy_length);
duke@435	5135
duke@435	5136	return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
duke@435	5137	}
duke@435	5138
duke@435	5139	// Helper function; generates the fast out-of-line call to an arraycopy stub.
duke@435	5140	void
duke@435	5141	LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
duke@435	5142	BasicType basic_elem_type,
duke@435	5143	bool disjoint_bases,
duke@435	5144	Node* src, Node* src_offset,
duke@435	5145	Node* dest, Node* dest_offset,
duke@435	5146	Node* copy_length) {
duke@435	5147	if (stopped()) return; // nothing to do
duke@435	5148
duke@435	5149	Node* src_start = src;
duke@435	5150	Node* dest_start = dest;
duke@435	5151	if (src_offset != NULL || dest_offset != NULL) {
duke@435	5152	assert(src_offset != NULL && dest_offset != NULL, "");
duke@435	5153	src_start = array_element_address(src, src_offset, basic_elem_type);
duke@435	5154	dest_start = array_element_address(dest, dest_offset, basic_elem_type);
duke@435	5155	}
duke@435	5156
duke@435	5157	// Figure out which arraycopy runtime method to call.
duke@435	5158	const char* copyfunc_name = "arraycopy";
duke@435	5159	address copyfunc_addr =
duke@435	5160	basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
duke@435	5161	disjoint_bases, copyfunc_name);
duke@435	5162
duke@435	5163	// Call it. Note that the count_ix value is not scaled to a byte-size.
duke@435	5164	make_runtime_call(RC_LEAF|RC_NO_FP,
duke@435	5165	OptoRuntime::fast_arraycopy_Type(),
duke@435	5166	copyfunc_addr, copyfunc_name, adr_type,
duke@435	5167	src_start, dest_start, copy_length XTOP);
duke@435	5168	}