4766 dest_spec = true; |
4688 dest_spec = true; |
4767 } |
4689 } |
4768 } |
4690 } |
4769 } |
4691 } |
4770 |
4692 |
4771 if (!has_src || !has_dest) { |
4693 if (has_src && has_dest) { |
4772 // Conservatively insert a memory barrier on all memory slices. |
4694 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); |
4773 // Do not let writes into the source float below the arraycopy. |
4695 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); |
4774 insert_mem_bar(Op_MemBarCPUOrder); |
4696 if (src_elem == T_ARRAY) src_elem = T_OBJECT; |
4775 |
4697 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; |
4776 // Call StubRoutines::generic_arraycopy stub. |
4698 |
4777 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, |
4699 if (src_elem == dest_elem && src_elem == T_OBJECT) { |
4778 src, src_offset, dest, dest_offset, length); |
4700 // If both arrays are object arrays then having the exact types |
4779 |
4701 // for both will remove the need for a subtype check at runtime |
4780 // Do not let reads from the destination float above the arraycopy. |
4702 // before the call and may make it possible to pick a faster copy |
4781 // Since we cannot type the arrays, we don't know which slices |
4703 // routine (without a subtype check on every element) |
4782 // might be affected. We could restrict this barrier only to those |
4704 // Do we have the exact type of src? |
4783 // memory slices which pertain to array elements--but don't bother. |
4705 bool could_have_src = src_spec; |
4784 if (!InsertMemBarAfterArraycopy) |
4706 // Do we have the exact type of dest? |
4785 // (If InsertMemBarAfterArraycopy, there is already one in place.) |
4707 bool could_have_dest = dest_spec; |
4786 insert_mem_bar(Op_MemBarCPUOrder); |
4708 ciKlass* src_k = top_src->klass(); |
4787 return true; |
4709 ciKlass* dest_k = top_dest->klass(); |
4788 } |
4710 if (!src_spec) { |
4789 |
4711 src_k = src_type->speculative_type_not_null(); |
4790 // (2) src and dest arrays must have elements of the same BasicType |
4712 if (src_k != NULL && src_k->is_array_klass()) { |
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; |
4713 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 } |
4714 } |
5101 } |
4715 } |
5102 if (didit) |
4716 if (!dest_spec) { |
5103 set_control(top()); // no regular fast path |
4717 dest_k = dest_type->speculative_type_not_null(); |
5104 } |
4718 if (dest_k != NULL && dest_k->is_array_klass()) { |
5105 |
4719 could_have_dest = true; |
5106 // Clear the tail, if any. |
4720 } |
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 } |
4721 } |
5128 } |
4722 if (could_have_src && could_have_dest) { |
5129 } |
4723 // If we can have both exact types, emit the missing guards |
5130 |
4724 if (could_have_src && !src_spec) { |
5131 BasicType copy_type = basic_elem_type; |
4725 src = maybe_cast_profiled_obj(src, src_k); |
5132 assert(basic_elem_type != T_ARRAY, "caller must fix this"); |
4726 } |
5133 if (!stopped() && copy_type == T_OBJECT) { |
4727 if (could_have_dest && !dest_spec) { |
5134 // If src and dest have compatible element types, we can copy bits. |
4728 dest = maybe_cast_profiled_obj(dest, dest_k); |
5135 // Types S[] and D[] are compatible if D is a supertype of S. |
4729 } |
5136 // |
4730 } |
5137 // If they are not, we will use checked_oop_disjoint_arraycopy, |
4731 } |
5138 // which performs a fast optimistic per-oop check, and backs off |
4732 } |
5139 // further to JVM_ArrayCopy on the first per-oop check that fails. |
4733 |
5140 // (Actually, we don't move raw bits only; the GC requires card marks.) |
4734 if (!too_many_traps(Deoptimization::Reason_intrinsic) && !src->is_top() && !dest->is_top()) { |
5141 |
4735 // validate arguments: enables transformation the ArrayCopyNode |
5142 // Get the Klass* for both src and dest |
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 |
5143 Node* src_klass = load_object_klass(src); |
4765 Node* src_klass = load_object_klass(src); |
5144 Node* dest_klass = load_object_klass(dest); |
4766 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); |
4767 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); |
5158 // Plug failing path into checked_oop_disjoint_arraycopy |
4768 |
5159 if (not_subtype_ctrl != top()) { |
4769 if (not_subtype_ctrl != top()) { |
5160 PreserveJVMState pjvms(this); |
4770 PreserveJVMState pjvms(this); |
5161 set_control(not_subtype_ctrl); |
4771 set_control(not_subtype_ctrl); |
5162 // (At this point we can assume disjoint_bases, since types differ.) |
4772 uncommon_trap(Deoptimization::Reason_intrinsic, |
5163 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); |
4773 Deoptimization::Action_make_not_entrant); |
5164 Node* p1 = basic_plus_adr(dest_klass, ek_offset); |
4774 assert(stopped(), "Should be stopped"); |
5165 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); |
4775 } |
5166 Node* dest_elem_klass = _gvn.transform(n1); |
4776 { |
5167 Node* cv = generate_checkcast_arraycopy(adr_type, |
4777 PreserveJVMState pjvms(this); |
5168 dest_elem_klass, |
4778 set_control(_gvn.transform(slow_region)); |
5169 src, src_offset, dest, dest_offset, |
4779 uncommon_trap(Deoptimization::Reason_intrinsic, |
5170 ConvI2X(copy_length), dest_uninitialized); |
4780 Deoptimization::Action_make_not_entrant); |
5171 if (cv == NULL) cv = intcon(-1); // failure (no stub available) |
4781 assert(stopped(), "Should be stopped"); |
5172 checked_control = control(); |
4782 } |
5173 checked_i_o = i_o(); |
4783 } |
5174 checked_mem = memory(adr_type); |
4784 |
5175 checked_value = cv; |
4785 if (stopped()) { |
5176 } |
4786 return true; |
5177 // At this point we know we do not need type checks on oop stores. |
4787 } |
5178 |
4788 |
5179 // Let's see if we need card marks: |
4789 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); |
5180 if (alloc != NULL && use_ReduceInitialCardMarks()) { |
4790 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL); |
5181 // If we do not need card marks, copy using the jint or jlong stub. |
4791 |
5182 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); |
4792 if (notest) { |
5183 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), |
4793 ac->set_arraycopy_notest(); |
5184 "sizes agree"); |
4794 } |
5185 } |
4795 |
5186 } |
4796 Node* n = _gvn.transform(ac); |
5187 |
4797 assert(n == ac, "cannot disappear"); |
5188 if (!stopped()) { |
4798 ac->connect_outputs(this); |
5189 // Generate the fast path, if possible. |
4799 |
5190 PreserveJVMState pjvms(this); |
4800 return true; |
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 } |
4801 } |
5322 |
4802 |
5323 |
4803 |
5324 // Helper function which determines if an arraycopy immediately follows |
4804 // Helper function which determines if an arraycopy immediately follows |
5325 // an allocation, with no intervening tests or other escapes for the object. |
4805 // an allocation, with no intervening tests or other escapes for the object. |
5395 // a new control state to which we will anchor the destination pointer. |
4875 // a new control state to which we will anchor the destination pointer. |
5396 |
4876 |
5397 return alloc; |
4877 return alloc; |
5398 } |
4878 } |
5399 |
4879 |
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----------------------------------- |
4880 //-------------inline_encodeISOArray----------------------------------- |
5700 // encode char[] to byte[] in ISO_8859_1 |
4881 // encode char[] to byte[] in ISO_8859_1 |
5701 bool LibraryCallKit::inline_encodeISOArray() { |
4882 bool LibraryCallKit::inline_encodeISOArray() { |
5702 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); |
4883 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); |
5703 // no receiver since it is static method |
4884 // no receiver since it is static method |