129 }
130
131 private:
132 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
133 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
134 }
135
136 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
137 void set_result(RegionNode* region, PhiNode* value);
138 Node* result() { return _result; }
139
140 virtual int reexecute_sp() { return _reexecute_sp; }
141
142 // Helper functions to inline natives
143 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
144 Node* generate_slow_guard(Node* test, RegionNode* region);
145 Node* generate_fair_guard(Node* test, RegionNode* region);
146 Node* generate_negative_guard(Node* index, RegionNode* region,
147 // resulting CastII of index:
148 Node* *pos_index = NULL);
149 Node* generate_nonpositive_guard(Node* index, bool never_negative,
150 // resulting CastII of index:
151 Node* *pos_index = NULL);
152 Node* generate_limit_guard(Node* offset, Node* subseq_length,
153 Node* array_length,
154 RegionNode* region);
155 Node* generate_current_thread(Node* &tls_output);
156 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
157 bool disjoint_bases, const char* &name, bool dest_uninitialized);
158 Node* load_mirror_from_klass(Node* klass);
159 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
160 RegionNode* region, int null_path,
161 int offset);
162 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
163 RegionNode* region, int null_path) {
164 int offset = java_lang_Class::klass_offset_in_bytes();
165 return load_klass_from_mirror_common(mirror, never_see_null,
166 region, null_path,
167 offset);
168 }
169 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
170 RegionNode* region, int null_path) {
171 int offset = java_lang_Class::array_klass_offset_in_bytes();
172 return load_klass_from_mirror_common(mirror, never_see_null,
173 region, null_path,
174 offset);
175 }
176 Node* generate_access_flags_guard(Node* kls,
177 int modifier_mask, int modifier_bits,
247 bool inline_native_threadID();
248 #endif
249 bool inline_native_time_funcs(address method, const char* funcName);
250 bool inline_native_isInterrupted();
251 bool inline_native_Class_query(vmIntrinsics::ID id);
252 bool inline_native_subtype_check();
253
254 bool inline_native_newArray();
255 bool inline_native_getLength();
256 bool inline_array_copyOf(bool is_copyOfRange);
257 bool inline_array_equals();
258 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
259 bool inline_native_clone(bool is_virtual);
260 bool inline_native_Reflection_getCallerClass();
261 // Helper function for inlining native object hash method
262 bool inline_native_hashcode(bool is_virtual, bool is_static);
263 bool inline_native_getClass();
264
265 // Helper functions for inlining arraycopy
266 bool inline_arraycopy();
267 void generate_arraycopy(const TypePtr* adr_type,
268 BasicType basic_elem_type,
269 Node* src, Node* src_offset,
270 Node* dest, Node* dest_offset,
271 Node* copy_length,
272 bool disjoint_bases = false,
273 bool length_never_negative = false,
274 RegionNode* slow_region = NULL);
275 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
276 RegionNode* slow_region);
277 void generate_clear_array(const TypePtr* adr_type,
278 Node* dest,
279 BasicType basic_elem_type,
280 Node* slice_off,
281 Node* slice_len,
282 Node* slice_end);
283 bool generate_block_arraycopy(const TypePtr* adr_type,
284 BasicType basic_elem_type,
285 AllocateNode* alloc,
286 Node* src, Node* src_offset,
287 Node* dest, Node* dest_offset,
288 Node* dest_size, bool dest_uninitialized);
289 void generate_slow_arraycopy(const TypePtr* adr_type,
290 Node* src, Node* src_offset,
291 Node* dest, Node* dest_offset,
292 Node* copy_length, bool dest_uninitialized);
293 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
294 Node* dest_elem_klass,
295 Node* src, Node* src_offset,
296 Node* dest, Node* dest_offset,
297 Node* copy_length, bool dest_uninitialized);
298 Node* generate_generic_arraycopy(const TypePtr* adr_type,
299 Node* src, Node* src_offset,
300 Node* dest, Node* dest_offset,
301 Node* copy_length, bool dest_uninitialized);
302 void generate_unchecked_arraycopy(const TypePtr* adr_type,
303 BasicType basic_elem_type,
304 bool disjoint_bases,
305 Node* src, Node* src_offset,
306 Node* dest, Node* dest_offset,
307 Node* copy_length, bool dest_uninitialized);
308 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
309 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
310 bool inline_unsafe_ordered_store(BasicType type);
311 bool inline_unsafe_fence(vmIntrinsics::ID id);
312 bool inline_fp_conversions(vmIntrinsics::ID id);
313 bool inline_number_methods(vmIntrinsics::ID id);
314 bool inline_reference_get();
315 bool inline_aescrypt_Block(vmIntrinsics::ID id);
316 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
317 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
318 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
319 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
320 bool inline_sha_implCompress(vmIntrinsics::ID id);
321 bool inline_digestBase_implCompressMB(int predicate);
322 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
323 bool long_state, address stubAddr, const char *stubName,
324 Node* src_start, Node* ofs, Node* limit);
325 Node* get_state_from_sha_object(Node *sha_object);
326 Node* get_state_from_sha5_object(Node *sha_object);
327 Node* inline_digestBase_implCompressMB_predicate(int predicate);
1032 }
1033
1034 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1035 Node* *pos_index) {
1036 if (stopped())
1037 return NULL; // already stopped
1038 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1039 return NULL; // index is already adequately typed
1040 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1041 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1042 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1043 if (is_neg != NULL && pos_index != NULL) {
1044 // Emulate effect of Parse::adjust_map_after_if.
1045 Node* ccast = new CastIINode(index, TypeInt::POS);
1046 ccast->set_req(0, control());
1047 (*pos_index) = _gvn.transform(ccast);
1048 }
1049 return is_neg;
1050 }
1051
1052 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1053 Node* *pos_index) {
1054 if (stopped())
1055 return NULL; // already stopped
1056 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1057 return NULL; // index is already adequately typed
1058 Node* cmp_le = _gvn.transform(new CmpINode(index, intcon(0)));
1059 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1060 Node* bol_le = _gvn.transform(new BoolNode(cmp_le, le_or_eq));
1061 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1062 if (is_notp != NULL && pos_index != NULL) {
1063 // Emulate effect of Parse::adjust_map_after_if.
1064 Node* ccast = new CastIINode(index, TypeInt::POS1);
1065 ccast->set_req(0, control());
1066 (*pos_index) = _gvn.transform(ccast);
1067 }
1068 return is_notp;
1069 }
1070
1071 // Make sure that 'position' is a valid limit index, in [0..length].
1072 // There are two equivalent plans for checking this:
1073 // A. (offset + copyLength) unsigned<= arrayLength
1074 // B. offset <= (arrayLength - copyLength)
1075 // We require that all of the values above, except for the sum and
1076 // difference, are already known to be non-negative.
1077 // Plan A is robust in the face of overflow, if offset and copyLength
1078 // are both hugely positive.
1079 //
1080 // Plan B is less direct and intuitive, but it does not overflow at
1081 // all, since the difference of two non-negatives is always
1082 // representable. Whenever Java methods must perform the equivalent
1083 // check they generally use Plan B instead of Plan A.
1084 // For the moment we use Plan A.
1085 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1086 Node* subseq_length,
1087 Node* array_length,
1088 RegionNode* region) {
1089 if (stopped())
1090 return NULL; // already stopped
3911 if (bailout->req() > 1) {
3912 PreserveJVMState pjvms(this);
3913 set_control(_gvn.transform(bailout));
3914 uncommon_trap(Deoptimization::Reason_intrinsic,
3915 Deoptimization::Action_maybe_recompile);
3916 }
3917
3918 if (!stopped()) {
3919 // How many elements will we copy from the original?
3920 // The answer is MinI(orig_length - start, length).
3921 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3922 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3923
3924 newcopy = new_array(klass_node, length, 0); // no argments to push
3925
3926 // Generate a direct call to the right arraycopy function(s).
3927 // We know the copy is disjoint but we might not know if the
3928 // oop stores need checking.
3929 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3930 // This will fail a store-check if x contains any non-nulls.
3931 bool disjoint_bases = true;
3932 // if start > orig_length then the length of the copy may be
3933 // negative.
3934 bool length_never_negative = !is_copyOfRange;
3935 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3936 original, start, newcopy, intcon(0), moved,
3937 disjoint_bases, length_never_negative);
3938 }
3939 } // original reexecute is set back here
3940
3941 C->set_has_split_ifs(true); // Has chance for split-if optimization
3942 if (!stopped()) {
3943 set_result(newcopy);
3944 }
3945 return true;
3946 }
3947
3948
3949 //----------------------generate_virtual_guard---------------------------
3950 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3951 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3952 RegionNode* slow_region) {
3953 ciMethod* method = callee();
3954 int vtable_index = method->vtable_index();
3955 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3956 err_msg_res("bad index %d", vtable_index));
3957 // Get the Method* out of the appropriate vtable entry.
4428 if (base_off % BytesPerLong != 0) {
4429 assert(UseCompressedClassPointers, "");
4430 if (is_array) {
4431 // Exclude length to copy by 8 bytes words.
4432 base_off += sizeof(int);
4433 } else {
4434 // Include klass to copy by 8 bytes words.
4435 base_off = instanceOopDesc::klass_offset_in_bytes();
4436 }
4437 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4438 }
4439 src = basic_plus_adr(src, base_off);
4440 dest = basic_plus_adr(dest, base_off);
4441
4442 // Compute the length also, if needed:
4443 Node* countx = size;
4444 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4445 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4446
4447 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4448 bool disjoint_bases = true;
4449 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4450 src, NULL, dest, NULL, countx,
4451 /*dest_uninitialized*/true);
4452
4453 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4454 if (card_mark) {
4455 assert(!is_array, "");
4456 // Put in store barrier for any and all oops we are sticking
4457 // into this object. (We could avoid this if we could prove
4458 // that the object type contains no oop fields at all.)
4459 Node* no_particular_value = NULL;
4460 Node* no_particular_field = NULL;
4461 int raw_adr_idx = Compile::AliasIdxRaw;
4462 post_barrier(control(),
4463 memory(raw_adr_type),
4464 alloc_obj,
4465 no_particular_field,
4466 raw_adr_idx,
4467 no_particular_value,
4468 T_OBJECT,
4469 false);
4470 }
4471
4540 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4541 int raw_adr_idx = Compile::AliasIdxRaw;
4542
4543 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4544 if (array_ctl != NULL) {
4545 // It's an array.
4546 PreserveJVMState pjvms(this);
4547 set_control(array_ctl);
4548 Node* obj_length = load_array_length(obj);
4549 Node* obj_size = NULL;
4550 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4551
4552 if (!use_ReduceInitialCardMarks()) {
4553 // If it is an oop array, it requires very special treatment,
4554 // because card marking is required on each card of the array.
4555 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4556 if (is_obja != NULL) {
4557 PreserveJVMState pjvms2(this);
4558 set_control(is_obja);
4559 // Generate a direct call to the right arraycopy function(s).
4560 bool disjoint_bases = true;
4561 bool length_never_negative = true;
4562 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4563 obj, intcon(0), alloc_obj, intcon(0),
4564 obj_length,
4565 disjoint_bases, length_never_negative);
4566 result_reg->init_req(_objArray_path, control());
4567 result_val->init_req(_objArray_path, alloc_obj);
4568 result_i_o ->set_req(_objArray_path, i_o());
4569 result_mem ->set_req(_objArray_path, reset_memory());
4570 }
4571 }
4572 // Otherwise, there are no card marks to worry about.
4573 // (We can dispense with card marks if we know the allocation
4574 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4575 // causes the non-eden paths to take compensating steps to
4576 // simulate a fresh allocation, so that no further
4577 // card marks are required in compiled code to initialize
4578 // the object.)
4579
4580 if (!stopped()) {
4581 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4582
4583 // Present the results of the copy.
4584 result_reg->init_req(_array_path, control());
4585 result_val->init_req(_array_path, alloc_obj);
4639 PreserveJVMState pjvms(this);
4640 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4641 Node* slow_result = set_results_for_java_call(slow_call);
4642 // this->control() comes from set_results_for_java_call
4643 result_reg->init_req(_slow_path, control());
4644 result_val->init_req(_slow_path, slow_result);
4645 result_i_o ->set_req(_slow_path, i_o());
4646 result_mem ->set_req(_slow_path, reset_memory());
4647 }
4648
4649 // Return the combined state.
4650 set_control( _gvn.transform(result_reg));
4651 set_i_o( _gvn.transform(result_i_o));
4652 set_all_memory( _gvn.transform(result_mem));
4653 } // original reexecute is set back here
4654
4655 set_result(_gvn.transform(result_val));
4656 return true;
4657 }
4658
4659 //------------------------------basictype2arraycopy----------------------------
4660 address LibraryCallKit::basictype2arraycopy(BasicType t,
4661 Node* src_offset,
4662 Node* dest_offset,
4663 bool disjoint_bases,
4664 const char* &name,
4665 bool dest_uninitialized) {
4666 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4667 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4668
4669 bool aligned = false;
4670 bool disjoint = disjoint_bases;
4671
4672 // if the offsets are the same, we can treat the memory regions as
4673 // disjoint, because either the memory regions are in different arrays,
4674 // or they are identical (which we can treat as disjoint.) We can also
4675 // treat a copy with a destination index less that the source index
4676 // as disjoint since a low->high copy will work correctly in this case.
4677 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4678 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4679 // both indices are constants
4680 int s_offs = src_offset_inttype->get_con();
4681 int d_offs = dest_offset_inttype->get_con();
4682 int element_size = type2aelembytes(t);
4683 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4684 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4685 if (s_offs >= d_offs) disjoint = true;
4686 } else if (src_offset == dest_offset && src_offset != NULL) {
4687 // This can occur if the offsets are identical non-constants.
4688 disjoint = true;
4689 }
4690
4691 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4692 }
4693
4694
4695 //------------------------------inline_arraycopy-----------------------
4696 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4697 // Object dest, int destPos,
4698 // int length);
4699 bool LibraryCallKit::inline_arraycopy() {
4700 // Get the arguments.
4701 Node* src = argument(0); // type: oop
4702 Node* src_offset = argument(1); // type: int
4703 Node* dest = argument(2); // type: oop
4704 Node* dest_offset = argument(3); // type: int
4705 Node* length = argument(4); // type: int
4706
4707 // Compile time checks. If any of these checks cannot be verified at compile time,
4708 // we do not make a fast path for this call. Instead, we let the call remain as it
4709 // is. The checks we choose to mandate at compile time are:
4710 //
4711 // (1) src and dest are arrays.
4712 const Type* src_type = src->Value(&_gvn);
4713 const Type* dest_type = dest->Value(&_gvn);
4714 const TypeAryPtr* top_src = src_type->isa_aryptr();
4715 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4716
4717 // Do we have the type of src?
4718 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4719 // Do we have the type of dest?
4720 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4721 // Is the type for src from speculation?
4722 bool src_spec = false;
4723 // Is the type for dest from speculation?
4724 bool dest_spec = false;
4725
4726 if (!has_src || !has_dest) {
4727 // We don't have sufficient type information, let's see if
4728 // speculative types can help. We need to have types for both src
4729 // and dest so that it pays off.
4730
4731 // Do we already have or could we have type information for src
4732 bool could_have_src = has_src;
4733 // Do we already have or could we have type information for dest
4751
4752 if (could_have_src && could_have_dest) {
4753 // This is going to pay off so emit the required guards
4754 if (!has_src) {
4755 src = maybe_cast_profiled_obj(src, src_k);
4756 src_type = _gvn.type(src);
4757 top_src = src_type->isa_aryptr();
4758 has_src = (top_src != NULL && top_src->klass() != NULL);
4759 src_spec = true;
4760 }
4761 if (!has_dest) {
4762 dest = maybe_cast_profiled_obj(dest, dest_k);
4763 dest_type = _gvn.type(dest);
4764 top_dest = dest_type->isa_aryptr();
4765 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4766 dest_spec = true;
4767 }
4768 }
4769 }
4770
4771 if (!has_src || !has_dest) {
4772 // Conservatively insert a memory barrier on all memory slices.
4773 // Do not let writes into the source float below the arraycopy.
4774 insert_mem_bar(Op_MemBarCPUOrder);
4775
4776 // Call StubRoutines::generic_arraycopy stub.
4777 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4778 src, src_offset, dest, dest_offset, length);
4779
4780 // Do not let reads from the destination float above the arraycopy.
4781 // Since we cannot type the arrays, we don't know which slices
4782 // might be affected. We could restrict this barrier only to those
4783 // memory slices which pertain to array elements--but don't bother.
4784 if (!InsertMemBarAfterArraycopy)
4785 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4786 insert_mem_bar(Op_MemBarCPUOrder);
4787 return true;
4788 }
4789
4790 // (2) src and dest arrays must have elements of the same BasicType
4791 // Figure out the size and type of the elements we will be copying.
4792 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4793 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4794 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4795 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4796
4797 if (src_elem != dest_elem || dest_elem == T_VOID) {
4798 // The component types are not the same or are not recognized. Punt.
4799 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4800 generate_slow_arraycopy(TypePtr::BOTTOM,
4801 src, src_offset, dest, dest_offset, length,
4802 /*dest_uninitialized*/false);
4803 return true;
4804 }
4805
4806 if (src_elem == T_OBJECT) {
4807 // If both arrays are object arrays then having the exact types
4808 // for both will remove the need for a subtype check at runtime
4809 // before the call and may make it possible to pick a faster copy
4810 // routine (without a subtype check on every element)
4811 // Do we have the exact type of src?
4812 bool could_have_src = src_spec;
4813 // Do we have the exact type of dest?
4814 bool could_have_dest = dest_spec;
4815 ciKlass* src_k = top_src->klass();
4816 ciKlass* dest_k = top_dest->klass();
4817 if (!src_spec) {
4818 src_k = src_type->speculative_type_not_null();
4819 if (src_k != NULL && src_k->is_array_klass()) {
4820 could_have_src = true;
4821 }
4822 }
4823 if (!dest_spec) {
4824 dest_k = dest_type->speculative_type_not_null();
4825 if (dest_k != NULL && dest_k->is_array_klass()) {
4826 could_have_dest = true;
4827 }
4828 }
4829 if (could_have_src && could_have_dest) {
4830 // If we can have both exact types, emit the missing guards
4831 if (could_have_src && !src_spec) {
4832 src = maybe_cast_profiled_obj(src, src_k);
4833 }
4834 if (could_have_dest && !dest_spec) {
4835 dest = maybe_cast_profiled_obj(dest, dest_k);
4836 }
4837 }
4838 }
4839
4840 //---------------------------------------------------------------------------
4841 // We will make a fast path for this call to arraycopy.
4842
4843 // We have the following tests left to perform:
4844 //
4845 // (3) src and dest must not be null.
4846 // (4) src_offset must not be negative.
4847 // (5) dest_offset must not be negative.
4848 // (6) length must not be negative.
4849 // (7) src_offset + length must not exceed length of src.
4850 // (8) dest_offset + length must not exceed length of dest.
4851 // (9) each element of an oop array must be assignable
4852
4853 RegionNode* slow_region = new RegionNode(1);
4854 record_for_igvn(slow_region);
4855
4856 // (3) operands must not be null
4857 // We currently perform our null checks with the null_check routine.
4858 // This means that the null exceptions will be reported in the caller
4859 // rather than (correctly) reported inside of the native arraycopy call.
4860 // This should be corrected, given time. We do our null check with the
4861 // stack pointer restored.
4862 src = null_check(src, T_ARRAY);
4863 dest = null_check(dest, T_ARRAY);
4864
4865 // (4) src_offset must not be negative.
4866 generate_negative_guard(src_offset, slow_region);
4867
4868 // (5) dest_offset must not be negative.
4869 generate_negative_guard(dest_offset, slow_region);
4870
4871 // (6) length must not be negative (moved to generate_arraycopy()).
4872 // generate_negative_guard(length, slow_region);
4873
4874 // (7) src_offset + length must not exceed length of src.
4875 generate_limit_guard(src_offset, length,
4876 load_array_length(src),
4877 slow_region);
4878
4879 // (8) dest_offset + length must not exceed length of dest.
4880 generate_limit_guard(dest_offset, length,
4881 load_array_length(dest),
4882 slow_region);
4883
4884 // (9) each element of an oop array must be assignable
4885 // The generate_arraycopy subroutine checks this.
4886
4887 // This is where the memory effects are placed:
4888 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4889 generate_arraycopy(adr_type, dest_elem,
4890 src, src_offset, dest, dest_offset, length,
4891 false, false, slow_region);
4892
4893 return true;
4894 }
4895
4896 //-----------------------------generate_arraycopy----------------------
4897 // Generate an optimized call to arraycopy.
4898 // Caller must guard against non-arrays.
4899 // Caller must determine a common array basic-type for both arrays.
4900 // Caller must validate offsets against array bounds.
4901 // The slow_region has already collected guard failure paths
4902 // (such as out of bounds length or non-conformable array types).
4903 // The generated code has this shape, in general:
4904 //
4905 // if (length == 0) return // via zero_path
4906 // slowval = -1
4907 // if (types unknown) {
4908 // slowval = call generic copy loop
4909 // if (slowval == 0) return // via checked_path
4910 // } else if (indexes in bounds) {
4911 // if ((is object array) && !(array type check)) {
4912 // slowval = call checked copy loop
4913 // if (slowval == 0) return // via checked_path
4914 // } else {
4915 // call bulk copy loop
4916 // return // via fast_path
4917 // }
4918 // }
4919 // // adjust params for remaining work:
4920 // if (slowval != -1) {
4921 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4922 // }
4923 // slow_region:
4924 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4925 // return // via slow_call_path
4926 //
4927 // This routine is used from several intrinsics: System.arraycopy,
4928 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4929 //
4930 void
4931 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4932 BasicType basic_elem_type,
4933 Node* src, Node* src_offset,
4934 Node* dest, Node* dest_offset,
4935 Node* copy_length,
4936 bool disjoint_bases,
4937 bool length_never_negative,
4938 RegionNode* slow_region) {
4939
4940 if (slow_region == NULL) {
4941 slow_region = new RegionNode(1);
4942 record_for_igvn(slow_region);
4943 }
4944
4945 Node* original_dest = dest;
4946 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4947 bool dest_uninitialized = false;
4948
4949 // See if this is the initialization of a newly-allocated array.
4950 // If so, we will take responsibility here for initializing it to zero.
4951 // (Note: Because tightly_coupled_allocation performs checks on the
4952 // out-edges of the dest, we need to avoid making derived pointers
4953 // from it until we have checked its uses.)
4954 if (ReduceBulkZeroing
4955 && !ZeroTLAB // pointless if already zeroed
4956 && basic_elem_type != T_CONFLICT // avoid corner case
4957 && !src->eqv_uncast(dest)
4958 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4959 != NULL)
4960 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4961 && alloc->maybe_set_complete(&_gvn)) {
4962 // "You break it, you buy it."
4963 InitializeNode* init = alloc->initialization();
4964 assert(init->is_complete(), "we just did this");
4965 init->set_complete_with_arraycopy();
4966 assert(dest->is_CheckCastPP(), "sanity");
4967 assert(dest->in(0)->in(0) == init, "dest pinned");
4968 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4969 // From this point on, every exit path is responsible for
4970 // initializing any non-copied parts of the object to zero.
4971 // Also, if this flag is set we make sure that arraycopy interacts properly
4972 // with G1, eliding pre-barriers. See CR 6627983.
4973 dest_uninitialized = true;
4974 } else {
4975 // No zeroing elimination here.
4976 alloc = NULL;
4977 //original_dest = dest;
4978 //dest_uninitialized = false;
4979 }
4980
4981 // Results are placed here:
4982 enum { fast_path = 1, // normal void-returning assembly stub
4983 checked_path = 2, // special assembly stub with cleanup
4984 slow_call_path = 3, // something went wrong; call the VM
4985 zero_path = 4, // bypass when length of copy is zero
4986 bcopy_path = 5, // copy primitive array by 64-bit blocks
4987 PATH_LIMIT = 6
4988 };
4989 RegionNode* result_region = new RegionNode(PATH_LIMIT);
4990 PhiNode* result_i_o = new PhiNode(result_region, Type::ABIO);
4991 PhiNode* result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
4992 record_for_igvn(result_region);
4993 _gvn.set_type_bottom(result_i_o);
4994 _gvn.set_type_bottom(result_memory);
4995 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4996
4997 // The slow_control path:
4998 Node* slow_control;
4999 Node* slow_i_o = i_o();
5000 Node* slow_mem = memory(adr_type);
5001 debug_only(slow_control = (Node*) badAddress);
5002
5003 // Checked control path:
5004 Node* checked_control = top();
5005 Node* checked_mem = NULL;
5006 Node* checked_i_o = NULL;
5007 Node* checked_value = NULL;
5008
5009 if (basic_elem_type == T_CONFLICT) {
5010 assert(!dest_uninitialized, "");
5011 Node* cv = generate_generic_arraycopy(adr_type,
5012 src, src_offset, dest, dest_offset,
5013 copy_length, dest_uninitialized);
5014 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5015 checked_control = control();
5016 checked_i_o = i_o();
5017 checked_mem = memory(adr_type);
5018 checked_value = cv;
5019 set_control(top()); // no fast path
5020 }
5021
5022 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5023 if (not_pos != NULL) {
5024 PreserveJVMState pjvms(this);
5025 set_control(not_pos);
5026
5027 // (6) length must not be negative.
5028 if (!length_never_negative) {
5029 generate_negative_guard(copy_length, slow_region);
5030 }
5031
5032 // copy_length is 0.
5033 if (!stopped() && dest_uninitialized) {
5034 Node* dest_length = alloc->in(AllocateNode::ALength);
5035 if (copy_length->eqv_uncast(dest_length)
5036 || _gvn.find_int_con(dest_length, 1) <= 0) {
5037 // There is no zeroing to do. No need for a secondary raw memory barrier.
5038 } else {
5039 // Clear the whole thing since there are no source elements to copy.
5040 generate_clear_array(adr_type, dest, basic_elem_type,
5041 intcon(0), NULL,
5042 alloc->in(AllocateNode::AllocSize));
5043 // Use a secondary InitializeNode as raw memory barrier.
5044 // Currently it is needed only on this path since other
5045 // paths have stub or runtime calls as raw memory barriers.
5046 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5047 Compile::AliasIdxRaw,
5048 top())->as_Initialize();
5049 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5050 }
5051 }
5052
5053 // Present the results of the fast call.
5054 result_region->init_req(zero_path, control());
5055 result_i_o ->init_req(zero_path, i_o());
5056 result_memory->init_req(zero_path, memory(adr_type));
5057 }
5058
5059 if (!stopped() && dest_uninitialized) {
5060 // We have to initialize the *uncopied* part of the array to zero.
5061 // The copy destination is the slice dest[off..off+len]. The other slices
5062 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5063 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5064 Node* dest_length = alloc->in(AllocateNode::ALength);
5065 Node* dest_tail = _gvn.transform(new AddINode(dest_offset, copy_length));
5066
5067 // If there is a head section that needs zeroing, do it now.
5068 if (find_int_con(dest_offset, -1) != 0) {
5069 generate_clear_array(adr_type, dest, basic_elem_type,
5070 intcon(0), dest_offset,
5071 NULL);
5072 }
5073
5074 // Next, perform a dynamic check on the tail length.
5075 // It is often zero, and we can win big if we prove this.
5076 // There are two wins: Avoid generating the ClearArray
5077 // with its attendant messy index arithmetic, and upgrade
5078 // the copy to a more hardware-friendly word size of 64 bits.
5079 Node* tail_ctl = NULL;
5080 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5081 Node* cmp_lt = _gvn.transform(new CmpINode(dest_tail, dest_length));
5082 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
5083 tail_ctl = generate_slow_guard(bol_lt, NULL);
5084 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5085 }
5086
5087 // At this point, let's assume there is no tail.
5088 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5089 // There is no tail. Try an upgrade to a 64-bit copy.
5090 bool didit = false;
5091 { PreserveJVMState pjvms(this);
5092 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5093 src, src_offset, dest, dest_offset,
5094 dest_size, dest_uninitialized);
5095 if (didit) {
5096 // Present the results of the block-copying fast call.
5097 result_region->init_req(bcopy_path, control());
5098 result_i_o ->init_req(bcopy_path, i_o());
5099 result_memory->init_req(bcopy_path, memory(adr_type));
5100 }
5101 }
5102 if (didit)
5103 set_control(top()); // no regular fast path
5104 }
5105
5106 // Clear the tail, if any.
5107 if (tail_ctl != NULL) {
5108 Node* notail_ctl = stopped() ? NULL : control();
5109 set_control(tail_ctl);
5110 if (notail_ctl == NULL) {
5111 generate_clear_array(adr_type, dest, basic_elem_type,
5112 dest_tail, NULL,
5113 dest_size);
5114 } else {
5115 // Make a local merge.
5116 Node* done_ctl = new RegionNode(3);
5117 Node* done_mem = new PhiNode(done_ctl, Type::MEMORY, adr_type);
5118 done_ctl->init_req(1, notail_ctl);
5119 done_mem->init_req(1, memory(adr_type));
5120 generate_clear_array(adr_type, dest, basic_elem_type,
5121 dest_tail, NULL,
5122 dest_size);
5123 done_ctl->init_req(2, control());
5124 done_mem->init_req(2, memory(adr_type));
5125 set_control( _gvn.transform(done_ctl));
5126 set_memory( _gvn.transform(done_mem), adr_type );
5127 }
5128 }
5129 }
5130
5131 BasicType copy_type = basic_elem_type;
5132 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5133 if (!stopped() && copy_type == T_OBJECT) {
5134 // If src and dest have compatible element types, we can copy bits.
5135 // Types S[] and D[] are compatible if D is a supertype of S.
5136 //
5137 // If they are not, we will use checked_oop_disjoint_arraycopy,
5138 // which performs a fast optimistic per-oop check, and backs off
5139 // further to JVM_ArrayCopy on the first per-oop check that fails.
5140 // (Actually, we don't move raw bits only; the GC requires card marks.)
5141
5142 // Get the Klass* for both src and dest
5143 Node* src_klass = load_object_klass(src);
5144 Node* dest_klass = load_object_klass(dest);
5145
5146 // Generate the subtype check.
5147 // This might fold up statically, or then again it might not.
5148 //
5149 // Non-static example: Copying List<String>.elements to a new String[].
5150 // The backing store for a List<String> is always an Object[],
5151 // but its elements are always type String, if the generic types
5152 // are correct at the source level.
5153 //
5154 // Test S[] against D[], not S against D, because (probably)
5155 // the secondary supertype cache is less busy for S[] than S.
5156 // This usually only matters when D is an interface.
5157 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5158 // Plug failing path into checked_oop_disjoint_arraycopy
5159 if (not_subtype_ctrl != top()) {
5160 PreserveJVMState pjvms(this);
5161 set_control(not_subtype_ctrl);
5162 // (At this point we can assume disjoint_bases, since types differ.)
5163 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5164 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5165 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5166 Node* dest_elem_klass = _gvn.transform(n1);
5167 Node* cv = generate_checkcast_arraycopy(adr_type,
5168 dest_elem_klass,
5169 src, src_offset, dest, dest_offset,
5170 ConvI2X(copy_length), dest_uninitialized);
5171 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5172 checked_control = control();
5173 checked_i_o = i_o();
5174 checked_mem = memory(adr_type);
5175 checked_value = cv;
5176 }
5177 // At this point we know we do not need type checks on oop stores.
5178
5179 // Let's see if we need card marks:
5180 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5181 // If we do not need card marks, copy using the jint or jlong stub.
5182 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5183 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5184 "sizes agree");
5185 }
5186 }
5187
5188 if (!stopped()) {
5189 // Generate the fast path, if possible.
5190 PreserveJVMState pjvms(this);
5191 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5192 src, src_offset, dest, dest_offset,
5193 ConvI2X(copy_length), dest_uninitialized);
5194
5195 // Present the results of the fast call.
5196 result_region->init_req(fast_path, control());
5197 result_i_o ->init_req(fast_path, i_o());
5198 result_memory->init_req(fast_path, memory(adr_type));
5199 }
5200
5201 // Here are all the slow paths up to this point, in one bundle:
5202 slow_control = top();
5203 if (slow_region != NULL)
5204 slow_control = _gvn.transform(slow_region);
5205 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5206
5207 set_control(checked_control);
5208 if (!stopped()) {
5209 // Clean up after the checked call.
5210 // The returned value is either 0 or -1^K,
5211 // where K = number of partially transferred array elements.
5212 Node* cmp = _gvn.transform(new CmpINode(checked_value, intcon(0)));
5213 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
5214 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5215
5216 // If it is 0, we are done, so transfer to the end.
5217 Node* checks_done = _gvn.transform(new IfTrueNode(iff));
5218 result_region->init_req(checked_path, checks_done);
5219 result_i_o ->init_req(checked_path, checked_i_o);
5220 result_memory->init_req(checked_path, checked_mem);
5221
5222 // If it is not zero, merge into the slow call.
5223 set_control( _gvn.transform(new IfFalseNode(iff) ));
5224 RegionNode* slow_reg2 = new RegionNode(3);
5225 PhiNode* slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
5226 PhiNode* slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
5227 record_for_igvn(slow_reg2);
5228 slow_reg2 ->init_req(1, slow_control);
5229 slow_i_o2 ->init_req(1, slow_i_o);
5230 slow_mem2 ->init_req(1, slow_mem);
5231 slow_reg2 ->init_req(2, control());
5232 slow_i_o2 ->init_req(2, checked_i_o);
5233 slow_mem2 ->init_req(2, checked_mem);
5234
5235 slow_control = _gvn.transform(slow_reg2);
5236 slow_i_o = _gvn.transform(slow_i_o2);
5237 slow_mem = _gvn.transform(slow_mem2);
5238
5239 if (alloc != NULL) {
5240 // We'll restart from the very beginning, after zeroing the whole thing.
5241 // This can cause double writes, but that's OK since dest is brand new.
5242 // So we ignore the low 31 bits of the value returned from the stub.
5243 } else {
5244 // We must continue the copy exactly where it failed, or else
5245 // another thread might see the wrong number of writes to dest.
5246 Node* checked_offset = _gvn.transform(new XorINode(checked_value, intcon(-1)));
5247 Node* slow_offset = new PhiNode(slow_reg2, TypeInt::INT);
5248 slow_offset->init_req(1, intcon(0));
5249 slow_offset->init_req(2, checked_offset);
5250 slow_offset = _gvn.transform(slow_offset);
5251
5252 // Adjust the arguments by the conditionally incoming offset.
5253 Node* src_off_plus = _gvn.transform(new AddINode(src_offset, slow_offset));
5254 Node* dest_off_plus = _gvn.transform(new AddINode(dest_offset, slow_offset));
5255 Node* length_minus = _gvn.transform(new SubINode(copy_length, slow_offset));
5256
5257 // Tweak the node variables to adjust the code produced below:
5258 src_offset = src_off_plus;
5259 dest_offset = dest_off_plus;
5260 copy_length = length_minus;
5261 }
5262 }
5263
5264 set_control(slow_control);
5265 if (!stopped()) {
5266 // Generate the slow path, if needed.
5267 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5268
5269 set_memory(slow_mem, adr_type);
5270 set_i_o(slow_i_o);
5271
5272 if (dest_uninitialized) {
5273 generate_clear_array(adr_type, dest, basic_elem_type,
5274 intcon(0), NULL,
5275 alloc->in(AllocateNode::AllocSize));
5276 }
5277
5278 generate_slow_arraycopy(adr_type,
5279 src, src_offset, dest, dest_offset,
5280 copy_length, /*dest_uninitialized*/false);
5281
5282 result_region->init_req(slow_call_path, control());
5283 result_i_o ->init_req(slow_call_path, i_o());
5284 result_memory->init_req(slow_call_path, memory(adr_type));
5285 }
5286
5287 // Remove unused edges.
5288 for (uint i = 1; i < result_region->req(); i++) {
5289 if (result_region->in(i) == NULL)
5290 result_region->init_req(i, top());
5291 }
5292
5293 // Finished; return the combined state.
5294 set_control( _gvn.transform(result_region));
5295 set_i_o( _gvn.transform(result_i_o) );
5296 set_memory( _gvn.transform(result_memory), adr_type );
5297
5298 // The memory edges above are precise in order to model effects around
5299 // array copies accurately to allow value numbering of field loads around
5300 // arraycopy. Such field loads, both before and after, are common in Java
5301 // collections and similar classes involving header/array data structures.
5302 //
5303 // But with low number of register or when some registers are used or killed
5304 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5305 // The next memory barrier is added to avoid it. If the arraycopy can be
5306 // optimized away (which it can, sometimes) then we can manually remove
5307 // the membar also.
5308 //
5309 // Do not let reads from the cloned object float above the arraycopy.
5310 if (alloc != NULL) {
5311 // Do not let stores that initialize this object be reordered with
5312 // a subsequent store that would make this object accessible by
5313 // other threads.
5314 // Record what AllocateNode this StoreStore protects so that
5315 // escape analysis can go from the MemBarStoreStoreNode to the
5316 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5317 // based on the escape status of the AllocateNode.
5318 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5319 } else if (InsertMemBarAfterArraycopy)
5320 insert_mem_bar(Op_MemBarCPUOrder);
5321 }
5322
5323
5324 // Helper function which determines if an arraycopy immediately follows
5325 // an allocation, with no intervening tests or other escapes for the object.
5326 AllocateArrayNode*
5327 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5328 RegionNode* slow_region) {
5329 if (stopped()) return NULL; // no fast path
5330 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5331
5332 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5333 if (alloc == NULL) return NULL;
5334
5335 Node* rawmem = memory(Compile::AliasIdxRaw);
5336 // Is the allocation's memory state untouched?
5337 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5338 // Bail out if there have been raw-memory effects since the allocation.
5339 // (Example: There might have been a call or safepoint.)
5340 return NULL;
5378 Node* obs = not_ctl->fast_out(j);
5379 if (obs->in(0) == not_ctl && obs->is_Call() &&
5380 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5381 found_trap = true; break;
5382 }
5383 }
5384 if (found_trap) {
5385 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5386 continue;
5387 }
5388 }
5389 return NULL;
5390 }
5391
5392 // If we get this far, we have an allocation which immediately
5393 // precedes the arraycopy, and we can take over zeroing the new object.
5394 // The arraycopy will finish the initialization, and provide
5395 // a new control state to which we will anchor the destination pointer.
5396
5397 return alloc;
5398 }
5399
5400 // Helper for initialization of arrays, creating a ClearArray.
5401 // It writes zero bits in [start..end), within the body of an array object.
5402 // The memory effects are all chained onto the 'adr_type' alias category.
5403 //
5404 // Since the object is otherwise uninitialized, we are free
5405 // to put a little "slop" around the edges of the cleared area,
5406 // as long as it does not go back into the array's header,
5407 // or beyond the array end within the heap.
5408 //
5409 // The lower edge can be rounded down to the nearest jint and the
5410 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5411 //
5412 // Arguments:
5413 // adr_type memory slice where writes are generated
5414 // dest oop of the destination array
5415 // basic_elem_type element type of the destination
5416 // slice_idx array index of first element to store
5417 // slice_len number of elements to store (or NULL)
5418 // dest_size total size in bytes of the array object
5419 //
5420 // Exactly one of slice_len or dest_size must be non-NULL.
5421 // If dest_size is non-NULL, zeroing extends to the end of the object.
5422 // If slice_len is non-NULL, the slice_idx value must be a constant.
5423 void
5424 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5425 Node* dest,
5426 BasicType basic_elem_type,
5427 Node* slice_idx,
5428 Node* slice_len,
5429 Node* dest_size) {
5430 // one or the other but not both of slice_len and dest_size:
5431 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5432 if (slice_len == NULL) slice_len = top();
5433 if (dest_size == NULL) dest_size = top();
5434
5435 // operate on this memory slice:
5436 Node* mem = memory(adr_type); // memory slice to operate on
5437
5438 // scaling and rounding of indexes:
5439 int scale = exact_log2(type2aelembytes(basic_elem_type));
5440 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5441 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5442 int bump_bit = (-1 << scale) & BytesPerInt;
5443
5444 // determine constant starts and ends
5445 const intptr_t BIG_NEG = -128;
5446 assert(BIG_NEG + 2*abase < 0, "neg enough");
5447 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5448 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5449 if (slice_len_con == 0) {
5450 return; // nothing to do here
5451 }
5452 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5453 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5454 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5455 assert(end_con < 0, "not two cons");
5456 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5457 BytesPerLong);
5458 }
5459
5460 if (start_con >= 0 && end_con >= 0) {
5461 // Constant start and end. Simple.
5462 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5463 start_con, end_con, &_gvn);
5464 } else if (start_con >= 0 && dest_size != top()) {
5465 // Constant start, pre-rounded end after the tail of the array.
5466 Node* end = dest_size;
5467 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5468 start_con, end, &_gvn);
5469 } else if (start_con >= 0 && slice_len != top()) {
5470 // Constant start, non-constant end. End needs rounding up.
5471 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5472 intptr_t end_base = abase + (slice_idx_con << scale);
5473 int end_round = (-1 << scale) & (BytesPerLong - 1);
5474 Node* end = ConvI2X(slice_len);
5475 if (scale != 0)
5476 end = _gvn.transform(new LShiftXNode(end, intcon(scale) ));
5477 end_base += end_round;
5478 end = _gvn.transform(new AddXNode(end, MakeConX(end_base)));
5479 end = _gvn.transform(new AndXNode(end, MakeConX(~end_round)));
5480 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5481 start_con, end, &_gvn);
5482 } else if (start_con < 0 && dest_size != top()) {
5483 // Non-constant start, pre-rounded end after the tail of the array.
5484 // This is almost certainly a "round-to-end" operation.
5485 Node* start = slice_idx;
5486 start = ConvI2X(start);
5487 if (scale != 0)
5488 start = _gvn.transform(new LShiftXNode( start, intcon(scale) ));
5489 start = _gvn.transform(new AddXNode(start, MakeConX(abase)));
5490 if ((bump_bit | clear_low) != 0) {
5491 int to_clear = (bump_bit | clear_low);
5492 // Align up mod 8, then store a jint zero unconditionally
5493 // just before the mod-8 boundary.
5494 if (((abase + bump_bit) & ~to_clear) - bump_bit
5495 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5496 bump_bit = 0;
5497 assert((abase & to_clear) == 0, "array base must be long-aligned");
5498 } else {
5499 // Bump 'start' up to (or past) the next jint boundary:
5500 start = _gvn.transform(new AddXNode(start, MakeConX(bump_bit)));
5501 assert((abase & clear_low) == 0, "array base must be int-aligned");
5502 }
5503 // Round bumped 'start' down to jlong boundary in body of array.
5504 start = _gvn.transform(new AndXNode(start, MakeConX(~to_clear)));
5505 if (bump_bit != 0) {
5506 // Store a zero to the immediately preceding jint:
5507 Node* x1 = _gvn.transform(new AddXNode(start, MakeConX(-bump_bit)));
5508 Node* p1 = basic_plus_adr(dest, x1);
5509 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5510 mem = _gvn.transform(mem);
5511 }
5512 }
5513 Node* end = dest_size; // pre-rounded
5514 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5515 start, end, &_gvn);
5516 } else {
5517 // Non-constant start, unrounded non-constant end.
5518 // (Nobody zeroes a random midsection of an array using this routine.)
5519 ShouldNotReachHere(); // fix caller
5520 }
5521
5522 // Done.
5523 set_memory(mem, adr_type);
5524 }
5525
5526
5527 bool
5528 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5529 BasicType basic_elem_type,
5530 AllocateNode* alloc,
5531 Node* src, Node* src_offset,
5532 Node* dest, Node* dest_offset,
5533 Node* dest_size, bool dest_uninitialized) {
5534 // See if there is an advantage from block transfer.
5535 int scale = exact_log2(type2aelembytes(basic_elem_type));
5536 if (scale >= LogBytesPerLong)
5537 return false; // it is already a block transfer
5538
5539 // Look at the alignment of the starting offsets.
5540 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5541
5542 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5543 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5544 if (src_off_con < 0 || dest_off_con < 0)
5545 // At present, we can only understand constants.
5546 return false;
5547
5548 intptr_t src_off = abase + (src_off_con << scale);
5549 intptr_t dest_off = abase + (dest_off_con << scale);
5550
5551 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5552 // Non-aligned; too bad.
5553 // One more chance: Pick off an initial 32-bit word.
5554 // This is a common case, since abase can be odd mod 8.
5555 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5556 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5557 Node* sptr = basic_plus_adr(src, src_off);
5558 Node* dptr = basic_plus_adr(dest, dest_off);
5559 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5560 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5561 src_off += BytesPerInt;
5562 dest_off += BytesPerInt;
5563 } else {
5564 return false;
5565 }
5566 }
5567 assert(src_off % BytesPerLong == 0, "");
5568 assert(dest_off % BytesPerLong == 0, "");
5569
5570 // Do this copy by giant steps.
5571 Node* sptr = basic_plus_adr(src, src_off);
5572 Node* dptr = basic_plus_adr(dest, dest_off);
5573 Node* countx = dest_size;
5574 countx = _gvn.transform(new SubXNode(countx, MakeConX(dest_off)));
5575 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong)));
5576
5577 bool disjoint_bases = true; // since alloc != NULL
5578 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5579 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5580
5581 return true;
5582 }
5583
5584
5585 // Helper function; generates code for the slow case.
5586 // We make a call to a runtime method which emulates the native method,
5587 // but without the native wrapper overhead.
5588 void
5589 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5590 Node* src, Node* src_offset,
5591 Node* dest, Node* dest_offset,
5592 Node* copy_length, bool dest_uninitialized) {
5593 assert(!dest_uninitialized, "Invariant");
5594 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5595 OptoRuntime::slow_arraycopy_Type(),
5596 OptoRuntime::slow_arraycopy_Java(),
5597 "slow_arraycopy", adr_type,
5598 src, src_offset, dest, dest_offset,
5599 copy_length);
5600
5601 // Handle exceptions thrown by this fellow:
5602 make_slow_call_ex(call, env()->Throwable_klass(), false);
5603 }
5604
5605 // Helper function; generates code for cases requiring runtime checks.
5606 Node*
5607 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5608 Node* dest_elem_klass,
5609 Node* src, Node* src_offset,
5610 Node* dest, Node* dest_offset,
5611 Node* copy_length, bool dest_uninitialized) {
5612 if (stopped()) return NULL;
5613
5614 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5615 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5616 return NULL;
5617 }
5618
5619 // Pick out the parameters required to perform a store-check
5620 // for the target array. This is an optimistic check. It will
5621 // look in each non-null element's class, at the desired klass's
5622 // super_check_offset, for the desired klass.
5623 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5624 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5625 Node* n3 = new LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5626 Node* check_offset = ConvI2X(_gvn.transform(n3));
5627 Node* check_value = dest_elem_klass;
5628
5629 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5630 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5631
5632 // (We know the arrays are never conjoint, because their types differ.)
5633 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5634 OptoRuntime::checkcast_arraycopy_Type(),
5635 copyfunc_addr, "checkcast_arraycopy", adr_type,
5636 // five arguments, of which two are
5637 // intptr_t (jlong in LP64)
5638 src_start, dest_start,
5639 copy_length XTOP,
5640 check_offset XTOP,
5641 check_value);
5642
5643 return _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5644 }
5645
5646
5647 // Helper function; generates code for cases requiring runtime checks.
5648 Node*
5649 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5650 Node* src, Node* src_offset,
5651 Node* dest, Node* dest_offset,
5652 Node* copy_length, bool dest_uninitialized) {
5653 assert(!dest_uninitialized, "Invariant");
5654 if (stopped()) return NULL;
5655 address copyfunc_addr = StubRoutines::generic_arraycopy();
5656 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5657 return NULL;
5658 }
5659
5660 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5661 OptoRuntime::generic_arraycopy_Type(),
5662 copyfunc_addr, "generic_arraycopy", adr_type,
5663 src, src_offset, dest, dest_offset, copy_length);
5664
5665 return _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5666 }
5667
5668 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5669 void
5670 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5671 BasicType basic_elem_type,
5672 bool disjoint_bases,
5673 Node* src, Node* src_offset,
5674 Node* dest, Node* dest_offset,
5675 Node* copy_length, bool dest_uninitialized) {
5676 if (stopped()) return; // nothing to do
5677
5678 Node* src_start = src;
5679 Node* dest_start = dest;
5680 if (src_offset != NULL || dest_offset != NULL) {
5681 assert(src_offset != NULL && dest_offset != NULL, "");
5682 src_start = array_element_address(src, src_offset, basic_elem_type);
5683 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5684 }
5685
5686 // Figure out which arraycopy runtime method to call.
5687 const char* copyfunc_name = "arraycopy";
5688 address copyfunc_addr =
5689 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5690 disjoint_bases, copyfunc_name, dest_uninitialized);
5691
5692 // Call it. Note that the count_ix value is not scaled to a byte-size.
5693 make_runtime_call(RC_LEAF|RC_NO_FP,
5694 OptoRuntime::fast_arraycopy_Type(),
5695 copyfunc_addr, copyfunc_name, adr_type,
5696 src_start, dest_start, copy_length XTOP);
5697 }
5698
5699 //-------------inline_encodeISOArray-----------------------------------
5700 // encode char[] to byte[] in ISO_8859_1
5701 bool LibraryCallKit::inline_encodeISOArray() {
5702 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5703 // no receiver since it is static method
5704 Node *src = argument(0);
5705 Node *src_offset = argument(1);
5706 Node *dst = argument(2);
5707 Node *dst_offset = argument(3);
5708 Node *length = argument(4);
5709
5710 const Type* src_type = src->Value(&_gvn);
5711 const Type* dst_type = dst->Value(&_gvn);
5712 const TypeAryPtr* top_src = src_type->isa_aryptr();
5713 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5714 if (top_src == NULL || top_src->klass() == NULL ||
5715 top_dest == NULL || top_dest->klass() == NULL) {
5716 // failed array check
|
129 }
130
131 private:
132 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
133 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
134 }
135
136 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
137 void set_result(RegionNode* region, PhiNode* value);
138 Node* result() { return _result; }
139
140 virtual int reexecute_sp() { return _reexecute_sp; }
141
142 // Helper functions to inline natives
143 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
144 Node* generate_slow_guard(Node* test, RegionNode* region);
145 Node* generate_fair_guard(Node* test, RegionNode* region);
146 Node* generate_negative_guard(Node* index, RegionNode* region,
147 // resulting CastII of index:
148 Node* *pos_index = NULL);
149 Node* generate_limit_guard(Node* offset, Node* subseq_length,
150 Node* array_length,
151 RegionNode* region);
152 Node* generate_current_thread(Node* &tls_output);
153 Node* load_mirror_from_klass(Node* klass);
154 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
155 RegionNode* region, int null_path,
156 int offset);
157 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
158 RegionNode* region, int null_path) {
159 int offset = java_lang_Class::klass_offset_in_bytes();
160 return load_klass_from_mirror_common(mirror, never_see_null,
161 region, null_path,
162 offset);
163 }
164 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
165 RegionNode* region, int null_path) {
166 int offset = java_lang_Class::array_klass_offset_in_bytes();
167 return load_klass_from_mirror_common(mirror, never_see_null,
168 region, null_path,
169 offset);
170 }
171 Node* generate_access_flags_guard(Node* kls,
172 int modifier_mask, int modifier_bits,
242 bool inline_native_threadID();
243 #endif
244 bool inline_native_time_funcs(address method, const char* funcName);
245 bool inline_native_isInterrupted();
246 bool inline_native_Class_query(vmIntrinsics::ID id);
247 bool inline_native_subtype_check();
248
249 bool inline_native_newArray();
250 bool inline_native_getLength();
251 bool inline_array_copyOf(bool is_copyOfRange);
252 bool inline_array_equals();
253 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
254 bool inline_native_clone(bool is_virtual);
255 bool inline_native_Reflection_getCallerClass();
256 // Helper function for inlining native object hash method
257 bool inline_native_hashcode(bool is_virtual, bool is_static);
258 bool inline_native_getClass();
259
260 // Helper functions for inlining arraycopy
261 bool inline_arraycopy();
262 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
263 RegionNode* slow_region);
264 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
265 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
266 bool inline_unsafe_ordered_store(BasicType type);
267 bool inline_unsafe_fence(vmIntrinsics::ID id);
268 bool inline_fp_conversions(vmIntrinsics::ID id);
269 bool inline_number_methods(vmIntrinsics::ID id);
270 bool inline_reference_get();
271 bool inline_aescrypt_Block(vmIntrinsics::ID id);
272 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
273 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
274 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
275 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
276 bool inline_sha_implCompress(vmIntrinsics::ID id);
277 bool inline_digestBase_implCompressMB(int predicate);
278 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
279 bool long_state, address stubAddr, const char *stubName,
280 Node* src_start, Node* ofs, Node* limit);
281 Node* get_state_from_sha_object(Node *sha_object);
282 Node* get_state_from_sha5_object(Node *sha_object);
283 Node* inline_digestBase_implCompressMB_predicate(int predicate);
988 }
989
990 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
991 Node* *pos_index) {
992 if (stopped())
993 return NULL; // already stopped
994 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
995 return NULL; // index is already adequately typed
996 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
997 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
998 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
999 if (is_neg != NULL && pos_index != NULL) {
1000 // Emulate effect of Parse::adjust_map_after_if.
1001 Node* ccast = new CastIINode(index, TypeInt::POS);
1002 ccast->set_req(0, control());
1003 (*pos_index) = _gvn.transform(ccast);
1004 }
1005 return is_neg;
1006 }
1007
1008 // Make sure that 'position' is a valid limit index, in [0..length].
1009 // There are two equivalent plans for checking this:
1010 // A. (offset + copyLength) unsigned<= arrayLength
1011 // B. offset <= (arrayLength - copyLength)
1012 // We require that all of the values above, except for the sum and
1013 // difference, are already known to be non-negative.
1014 // Plan A is robust in the face of overflow, if offset and copyLength
1015 // are both hugely positive.
1016 //
1017 // Plan B is less direct and intuitive, but it does not overflow at
1018 // all, since the difference of two non-negatives is always
1019 // representable. Whenever Java methods must perform the equivalent
1020 // check they generally use Plan B instead of Plan A.
1021 // For the moment we use Plan A.
1022 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1023 Node* subseq_length,
1024 Node* array_length,
1025 RegionNode* region) {
1026 if (stopped())
1027 return NULL; // already stopped
3848 if (bailout->req() > 1) {
3849 PreserveJVMState pjvms(this);
3850 set_control(_gvn.transform(bailout));
3851 uncommon_trap(Deoptimization::Reason_intrinsic,
3852 Deoptimization::Action_maybe_recompile);
3853 }
3854
3855 if (!stopped()) {
3856 // How many elements will we copy from the original?
3857 // The answer is MinI(orig_length - start, length).
3858 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3859 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3860
3861 newcopy = new_array(klass_node, length, 0); // no argments to push
3862
3863 // Generate a direct call to the right arraycopy function(s).
3864 // We know the copy is disjoint but we might not know if the
3865 // oop stores need checking.
3866 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3867 // This will fail a store-check if x contains any non-nulls.
3868
3869 Node* alloc = tightly_coupled_allocation(newcopy, NULL);
3870
3871 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, alloc != NULL);
3872 if (!is_copyOfRange) {
3873 ac->set_copyof();
3874 } else {
3875 ac->set_copyofrange();
3876 }
3877 Node* n = _gvn.transform(ac);
3878 assert(n == ac, "cannot disappear");
3879 ac->connect_outputs(this);
3880 }
3881 } // original reexecute is set back here
3882
3883 C->set_has_split_ifs(true); // Has chance for split-if optimization
3884 if (!stopped()) {
3885 set_result(newcopy);
3886 }
3887 return true;
3888 }
3889
3890
3891 //----------------------generate_virtual_guard---------------------------
3892 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3893 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3894 RegionNode* slow_region) {
3895 ciMethod* method = callee();
3896 int vtable_index = method->vtable_index();
3897 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3898 err_msg_res("bad index %d", vtable_index));
3899 // Get the Method* out of the appropriate vtable entry.
4370 if (base_off % BytesPerLong != 0) {
4371 assert(UseCompressedClassPointers, "");
4372 if (is_array) {
4373 // Exclude length to copy by 8 bytes words.
4374 base_off += sizeof(int);
4375 } else {
4376 // Include klass to copy by 8 bytes words.
4377 base_off = instanceOopDesc::klass_offset_in_bytes();
4378 }
4379 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4380 }
4381 src = basic_plus_adr(src, base_off);
4382 dest = basic_plus_adr(dest, base_off);
4383
4384 // Compute the length also, if needed:
4385 Node* countx = size;
4386 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4387 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4388
4389 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4390
4391 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4392 ac->set_clonebasic();
4393 Node* n = _gvn.transform(ac);
4394 assert(n == ac, "cannot disappear");
4395 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4396
4397 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4398 if (card_mark) {
4399 assert(!is_array, "");
4400 // Put in store barrier for any and all oops we are sticking
4401 // into this object. (We could avoid this if we could prove
4402 // that the object type contains no oop fields at all.)
4403 Node* no_particular_value = NULL;
4404 Node* no_particular_field = NULL;
4405 int raw_adr_idx = Compile::AliasIdxRaw;
4406 post_barrier(control(),
4407 memory(raw_adr_type),
4408 alloc_obj,
4409 no_particular_field,
4410 raw_adr_idx,
4411 no_particular_value,
4412 T_OBJECT,
4413 false);
4414 }
4415
4484 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4485 int raw_adr_idx = Compile::AliasIdxRaw;
4486
4487 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4488 if (array_ctl != NULL) {
4489 // It's an array.
4490 PreserveJVMState pjvms(this);
4491 set_control(array_ctl);
4492 Node* obj_length = load_array_length(obj);
4493 Node* obj_size = NULL;
4494 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4495
4496 if (!use_ReduceInitialCardMarks()) {
4497 // If it is an oop array, it requires very special treatment,
4498 // because card marking is required on each card of the array.
4499 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4500 if (is_obja != NULL) {
4501 PreserveJVMState pjvms2(this);
4502 set_control(is_obja);
4503 // Generate a direct call to the right arraycopy function(s).
4504 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4505 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4506 ac->set_cloneoop();
4507 Node* n = _gvn.transform(ac);
4508 assert(n == ac, "cannot disappear");
4509 ac->connect_outputs(this);
4510
4511 result_reg->init_req(_objArray_path, control());
4512 result_val->init_req(_objArray_path, alloc_obj);
4513 result_i_o ->set_req(_objArray_path, i_o());
4514 result_mem ->set_req(_objArray_path, reset_memory());
4515 }
4516 }
4517 // Otherwise, there are no card marks to worry about.
4518 // (We can dispense with card marks if we know the allocation
4519 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4520 // causes the non-eden paths to take compensating steps to
4521 // simulate a fresh allocation, so that no further
4522 // card marks are required in compiled code to initialize
4523 // the object.)
4524
4525 if (!stopped()) {
4526 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4527
4528 // Present the results of the copy.
4529 result_reg->init_req(_array_path, control());
4530 result_val->init_req(_array_path, alloc_obj);
4584 PreserveJVMState pjvms(this);
4585 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4586 Node* slow_result = set_results_for_java_call(slow_call);
4587 // this->control() comes from set_results_for_java_call
4588 result_reg->init_req(_slow_path, control());
4589 result_val->init_req(_slow_path, slow_result);
4590 result_i_o ->set_req(_slow_path, i_o());
4591 result_mem ->set_req(_slow_path, reset_memory());
4592 }
4593
4594 // Return the combined state.
4595 set_control( _gvn.transform(result_reg));
4596 set_i_o( _gvn.transform(result_i_o));
4597 set_all_memory( _gvn.transform(result_mem));
4598 } // original reexecute is set back here
4599
4600 set_result(_gvn.transform(result_val));
4601 return true;
4602 }
4603
4604 //------------------------------inline_arraycopy-----------------------
4605 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4606 // Object dest, int destPos,
4607 // int length);
4608 bool LibraryCallKit::inline_arraycopy() {
4609 // Get the arguments.
4610 Node* src = argument(0); // type: oop
4611 Node* src_offset = argument(1); // type: int
4612 Node* dest = argument(2); // type: oop
4613 Node* dest_offset = argument(3); // type: int
4614 Node* length = argument(4); // type: int
4615
4616 // The following tests must be performed
4617 // (1) src and dest are arrays.
4618 // (2) src and dest arrays must have elements of the same BasicType
4619 // (3) src and dest must not be null.
4620 // (4) src_offset must not be negative.
4621 // (5) dest_offset must not be negative.
4622 // (6) length must not be negative.
4623 // (7) src_offset + length must not exceed length of src.
4624 // (8) dest_offset + length must not exceed length of dest.
4625 // (9) each element of an oop array must be assignable
4626
4627 // (3) src and dest must not be null.
4628 // always do this here because we need the JVM state for uncommon traps
4629 src = null_check(src, T_ARRAY);
4630 dest = null_check(dest, T_ARRAY);
4631
4632 bool notest = false;
4633
4634 const Type* src_type = _gvn.type(src);
4635 const Type* dest_type = _gvn.type(dest);
4636 const TypeAryPtr* top_src = src_type->isa_aryptr();
4637 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4638
4639 // Do we have the type of src?
4640 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4641 // Do we have the type of dest?
4642 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4643 // Is the type for src from speculation?
4644 bool src_spec = false;
4645 // Is the type for dest from speculation?
4646 bool dest_spec = false;
4647
4648 if (!has_src || !has_dest) {
4649 // We don't have sufficient type information, let's see if
4650 // speculative types can help. We need to have types for both src
4651 // and dest so that it pays off.
4652
4653 // Do we already have or could we have type information for src
4654 bool could_have_src = has_src;
4655 // Do we already have or could we have type information for dest
4673
4674 if (could_have_src && could_have_dest) {
4675 // This is going to pay off so emit the required guards
4676 if (!has_src) {
4677 src = maybe_cast_profiled_obj(src, src_k);
4678 src_type = _gvn.type(src);
4679 top_src = src_type->isa_aryptr();
4680 has_src = (top_src != NULL && top_src->klass() != NULL);
4681 src_spec = true;
4682 }
4683 if (!has_dest) {
4684 dest = maybe_cast_profiled_obj(dest, dest_k);
4685 dest_type = _gvn.type(dest);
4686 top_dest = dest_type->isa_aryptr();
4687 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4688 dest_spec = true;
4689 }
4690 }
4691 }
4692
4693 if (has_src && has_dest) {
4694 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4695 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4696 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4697 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4698
4699 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4700 // If both arrays are object arrays then having the exact types
4701 // for both will remove the need for a subtype check at runtime
4702 // before the call and may make it possible to pick a faster copy
4703 // routine (without a subtype check on every element)
4704 // Do we have the exact type of src?
4705 bool could_have_src = src_spec;
4706 // Do we have the exact type of dest?
4707 bool could_have_dest = dest_spec;
4708 ciKlass* src_k = top_src->klass();
4709 ciKlass* dest_k = top_dest->klass();
4710 if (!src_spec) {
4711 src_k = src_type->speculative_type_not_null();
4712 if (src_k != NULL && src_k->is_array_klass()) {
4713 could_have_src = true;
4714 }
4715 }
4716 if (!dest_spec) {
4717 dest_k = dest_type->speculative_type_not_null();
4718 if (dest_k != NULL && dest_k->is_array_klass()) {
4719 could_have_dest = true;
4720 }
4721 }
4722 if (could_have_src && could_have_dest) {
4723 // If we can have both exact types, emit the missing guards
4724 if (could_have_src && !src_spec) {
4725 src = maybe_cast_profiled_obj(src, src_k);
4726 }
4727 if (could_have_dest && !dest_spec) {
4728 dest = maybe_cast_profiled_obj(dest, dest_k);
4729 }
4730 }
4731 }
4732 }
4733
4734 if (!too_many_traps(Deoptimization::Reason_intrinsic) && !src->is_top() && !dest->is_top()) {
4735 // validate arguments: enables transformation the ArrayCopyNode
4736 notest = true;
4737
4738 RegionNode* slow_region = new RegionNode(1);
4739 record_for_igvn(slow_region);
4740
4741 // (1) src and dest are arrays.
4742 generate_non_array_guard(load_object_klass(src), slow_region);
4743 generate_non_array_guard(load_object_klass(dest), slow_region);
4744
4745 // (2) src and dest arrays must have elements of the same BasicType
4746 // done at macro expansion or at Ideal transformation time
4747
4748 // (4) src_offset must not be negative.
4749 generate_negative_guard(src_offset, slow_region);
4750
4751 // (5) dest_offset must not be negative.
4752 generate_negative_guard(dest_offset, slow_region);
4753
4754 // (7) src_offset + length must not exceed length of src.
4755 generate_limit_guard(src_offset, length,
4756 load_array_length(src),
4757 slow_region);
4758
4759 // (8) dest_offset + length must not exceed length of dest.
4760 generate_limit_guard(dest_offset, length,
4761 load_array_length(dest),
4762 slow_region);
4763
4764 // (9) each element of an oop array must be assignable
4765 Node* src_klass = load_object_klass(src);
4766 Node* dest_klass = load_object_klass(dest);
4767 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4768
4769 if (not_subtype_ctrl != top()) {
4770 PreserveJVMState pjvms(this);
4771 set_control(not_subtype_ctrl);
4772 uncommon_trap(Deoptimization::Reason_intrinsic,
4773 Deoptimization::Action_make_not_entrant);
4774 assert(stopped(), "Should be stopped");
4775 }
4776 {
4777 PreserveJVMState pjvms(this);
4778 set_control(_gvn.transform(slow_region));
4779 uncommon_trap(Deoptimization::Reason_intrinsic,
4780 Deoptimization::Action_make_not_entrant);
4781 assert(stopped(), "Should be stopped");
4782 }
4783 }
4784
4785 if (stopped()) {
4786 return true;
4787 }
4788
4789 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4790 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL);
4791
4792 if (notest) {
4793 ac->set_arraycopy_notest();
4794 }
4795
4796 Node* n = _gvn.transform(ac);
4797 assert(n == ac, "cannot disappear");
4798 ac->connect_outputs(this);
4799
4800 return true;
4801 }
4802
4803
4804 // Helper function which determines if an arraycopy immediately follows
4805 // an allocation, with no intervening tests or other escapes for the object.
4806 AllocateArrayNode*
4807 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4808 RegionNode* slow_region) {
4809 if (stopped()) return NULL; // no fast path
4810 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4811
4812 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4813 if (alloc == NULL) return NULL;
4814
4815 Node* rawmem = memory(Compile::AliasIdxRaw);
4816 // Is the allocation's memory state untouched?
4817 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4818 // Bail out if there have been raw-memory effects since the allocation.
4819 // (Example: There might have been a call or safepoint.)
4820 return NULL;
4858 Node* obs = not_ctl->fast_out(j);
4859 if (obs->in(0) == not_ctl && obs->is_Call() &&
4860 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4861 found_trap = true; break;
4862 }
4863 }
4864 if (found_trap) {
4865 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4866 continue;
4867 }
4868 }
4869 return NULL;
4870 }
4871
4872 // If we get this far, we have an allocation which immediately
4873 // precedes the arraycopy, and we can take over zeroing the new object.
4874 // The arraycopy will finish the initialization, and provide
4875 // a new control state to which we will anchor the destination pointer.
4876
4877 return alloc;
4878 }
4879
4880 //-------------inline_encodeISOArray-----------------------------------
4881 // encode char[] to byte[] in ISO_8859_1
4882 bool LibraryCallKit::inline_encodeISOArray() {
4883 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4884 // no receiver since it is static method
4885 Node *src = argument(0);
4886 Node *src_offset = argument(1);
4887 Node *dst = argument(2);
4888 Node *dst_offset = argument(3);
4889 Node *length = argument(4);
4890
4891 const Type* src_type = src->Value(&_gvn);
4892 const Type* dst_type = dst->Value(&_gvn);
4893 const TypeAryPtr* top_src = src_type->isa_aryptr();
4894 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
4895 if (top_src == NULL || top_src->klass() == NULL ||
4896 top_dest == NULL || top_dest->klass() == NULL) {
4897 // failed array check
|