hotspot/src/share/vm/opto/arraycopynode.cpp
author roland
Wed, 18 Feb 2015 18:14:07 +0100
changeset 29337 ef2be52deeaf
child 29340 e5ee51840b93
permissions -rw-r--r--
6912521: System.arraycopy works slower than the simple loop for little lengths Summary: convert small array copies to series of loads and stores Reviewed-by: kvn, vlivanov
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
29337
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
     1
/*
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
     2
 * Copyright (c) 2015, Oracle and/or its affiliates. All rights reserved.
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
     3
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
     4
 *
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
     5
 * This code is free software; you can redistribute it and/or modify it
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
     6
 * under the terms of the GNU General Public License version 2 only, as
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
     7
 * published by the Free Software Foundation.
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
     8
 *
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
     9
 * This code is distributed in the hope that it will be useful, but WITHOUT
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    10
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    11
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    12
 * version 2 for more details (a copy is included in the LICENSE file that
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    13
 * accompanied this code).
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    14
 *
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    15
 * You should have received a copy of the GNU General Public License version
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    16
 * 2 along with this work; if not, write to the Free Software Foundation,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    17
 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    18
 *
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    19
 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    20
 * or visit www.oracle.com if you need additional information or have any
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    21
 * questions.
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    22
 *
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    23
 */
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    24
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    25
#include "precompiled.hpp"
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    26
#include "opto/arraycopynode.hpp"
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    27
#include "opto/graphKit.hpp"
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    28
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    29
ArrayCopyNode::ArrayCopyNode(Compile* C, bool alloc_tightly_coupled)
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    30
  : CallNode(arraycopy_type(), NULL, TypeRawPtr::BOTTOM),
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    31
    _alloc_tightly_coupled(alloc_tightly_coupled),
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    32
    _kind(None),
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    33
    _arguments_validated(false) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    34
  init_class_id(Class_ArrayCopy);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    35
  init_flags(Flag_is_macro);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    36
  C->add_macro_node(this);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    37
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    38
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    39
uint ArrayCopyNode::size_of() const { return sizeof(*this); }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    40
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    41
ArrayCopyNode* ArrayCopyNode::make(GraphKit* kit, bool may_throw,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    42
                                   Node* src, Node* src_offset,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    43
                                   Node* dest, Node* dest_offset,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    44
                                   Node* length,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    45
                                   bool alloc_tightly_coupled,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    46
                                   Node* src_klass, Node* dest_klass,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    47
                                   Node* src_length, Node* dest_length) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    48
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    49
  ArrayCopyNode* ac = new ArrayCopyNode(kit->C, alloc_tightly_coupled);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    50
  Node* prev_mem = kit->set_predefined_input_for_runtime_call(ac);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    51
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    52
  ac->init_req(ArrayCopyNode::Src, src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    53
  ac->init_req(ArrayCopyNode::SrcPos, src_offset);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    54
  ac->init_req(ArrayCopyNode::Dest, dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    55
  ac->init_req(ArrayCopyNode::DestPos, dest_offset);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    56
  ac->init_req(ArrayCopyNode::Length, length);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    57
  ac->init_req(ArrayCopyNode::SrcLen, src_length);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    58
  ac->init_req(ArrayCopyNode::DestLen, dest_length);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    59
  ac->init_req(ArrayCopyNode::SrcKlass, src_klass);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    60
  ac->init_req(ArrayCopyNode::DestKlass, dest_klass);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    61
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    62
  if (may_throw) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    63
    ac->set_req(TypeFunc::I_O , kit->i_o());
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    64
    kit->add_safepoint_edges(ac, false);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    65
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    66
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    67
  return ac;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    68
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    69
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    70
void ArrayCopyNode::connect_outputs(GraphKit* kit) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    71
  kit->set_all_memory_call(this, true);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    72
  kit->set_control(kit->gvn().transform(new ProjNode(this,TypeFunc::Control)));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    73
  kit->set_i_o(kit->gvn().transform(new ProjNode(this, TypeFunc::I_O)));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    74
  kit->make_slow_call_ex(this, kit->env()->Throwable_klass(), true);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    75
  kit->set_all_memory_call(this);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    76
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    77
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    78
#ifndef PRODUCT
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    79
const char* ArrayCopyNode::_kind_names[] = {"arraycopy", "arraycopy, validated arguments", "clone", "oop array clone", "CopyOf", "CopyOfRange"};
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    80
void ArrayCopyNode::dump_spec(outputStream *st) const {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    81
  CallNode::dump_spec(st);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    82
  st->print(" (%s%s)", _kind_names[_kind], _alloc_tightly_coupled ? ", tightly coupled allocation" : "");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    83
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    84
#endif
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    85
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    86
intptr_t ArrayCopyNode::get_length_if_constant(PhaseGVN *phase) const {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    87
  // check that length is constant
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    88
  Node* length = in(ArrayCopyNode::Length);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    89
  const Type* length_type = phase->type(length);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    90
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    91
  if (length_type == Type::TOP) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    92
    return -1;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    93
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    94
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    95
  assert(is_clonebasic() || is_arraycopy() || is_copyof() || is_copyofrange(), "unexpected array copy type");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    96
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    97
  return is_clonebasic() ? length->find_intptr_t_con(-1) : length->find_int_con(-1);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    98
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
    99
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   100
int ArrayCopyNode::get_count(PhaseGVN *phase) const {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   101
  Node* src = in(ArrayCopyNode::Src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   102
  const Type* src_type = phase->type(src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   103
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   104
  if (is_clonebasic()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   105
    if (src_type->isa_instptr()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   106
      const TypeInstPtr* inst_src = src_type->is_instptr();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   107
      ciInstanceKlass* ik = inst_src->klass()->as_instance_klass();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   108
      // ciInstanceKlass::nof_nonstatic_fields() doesn't take injected
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   109
      // fields into account. They are rare anyway so easier to simply
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   110
      // skip instances with injected fields.
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   111
      if ((!inst_src->klass_is_exact() && (ik->is_interface() || ik->has_subklass())) || ik->has_injected_fields()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   112
        return -1;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   113
      }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   114
      int nb_fields = ik->nof_nonstatic_fields();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   115
      return nb_fields;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   116
    } else {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   117
      const TypeAryPtr* ary_src = src_type->isa_aryptr();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   118
      assert (ary_src != NULL, "not an array or instance?");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   119
      // clone passes a length as a rounded number of longs. If we're
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   120
      // cloning an array we'll do it element by element. If the
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   121
      // length input to ArrayCopyNode is constant, length of input
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   122
      // array must be too.
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   123
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   124
      assert((get_length_if_constant(phase) == -1) == !ary_src->size()->is_con(), "inconsistent");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   125
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   126
      if (ary_src->size()->is_con()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   127
        return ary_src->size()->get_con();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   128
      }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   129
      return -1;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   130
    }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   131
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   132
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   133
  return get_length_if_constant(phase);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   134
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   135
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   136
Node* ArrayCopyNode::try_clone_instance(PhaseGVN *phase, bool can_reshape, int count) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   137
  if (!is_clonebasic()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   138
    return NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   139
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   140
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   141
  Node* src = in(ArrayCopyNode::Src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   142
  Node* dest = in(ArrayCopyNode::Dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   143
  Node* ctl = in(TypeFunc::Control);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   144
  Node* in_mem = in(TypeFunc::Memory);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   145
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   146
  const Type* src_type = phase->type(src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   147
  const Type* dest_type = phase->type(dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   148
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   149
  assert(src->is_AddP(), "should be base + off");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   150
  assert(dest->is_AddP(), "should be base + off");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   151
  Node* base_src = src->in(AddPNode::Base);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   152
  Node* base_dest = dest->in(AddPNode::Base);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   153
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   154
  MergeMemNode* mem = MergeMemNode::make(in_mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   155
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   156
  const TypeInstPtr* inst_src = src_type->isa_instptr();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   157
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   158
  if (inst_src == NULL) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   159
    return NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   160
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   161
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   162
  if (!inst_src->klass_is_exact()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   163
    ciInstanceKlass* ik = inst_src->klass()->as_instance_klass();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   164
    assert(!ik->is_interface() && !ik->has_subklass(), "inconsistent klass hierarchy");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   165
    phase->C->dependencies()->assert_leaf_type(ik);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   166
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   167
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   168
  ciInstanceKlass* ik = inst_src->klass()->as_instance_klass();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   169
  assert(ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem, "too many fields");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   170
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   171
  for (int i = 0; i < count; i++) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   172
    ciField* field = ik->nonstatic_field_at(i);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   173
    int fieldidx = phase->C->alias_type(field)->index();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   174
    const TypePtr* adr_type = phase->C->alias_type(field)->adr_type();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   175
    Node* off = phase->MakeConX(field->offset());
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   176
    Node* next_src = phase->transform(new AddPNode(base_src,base_src,off));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   177
    Node* next_dest = phase->transform(new AddPNode(base_dest,base_dest,off));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   178
    BasicType bt = field->layout_type();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   179
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   180
    const Type *type;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   181
    if (bt == T_OBJECT) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   182
      if (!field->type()->is_loaded()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   183
        type = TypeInstPtr::BOTTOM;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   184
      } else {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   185
        ciType* field_klass = field->type();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   186
        type = TypeOopPtr::make_from_klass(field_klass->as_klass());
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   187
      }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   188
    } else {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   189
      type = Type::get_const_basic_type(bt);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   190
    }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   191
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   192
    Node* v = LoadNode::make(*phase, ctl, mem->memory_at(fieldidx), next_src, adr_type, type, bt, MemNode::unordered);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   193
    v = phase->transform(v);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   194
    Node* s = StoreNode::make(*phase, ctl, mem->memory_at(fieldidx), next_dest, adr_type, v, bt, MemNode::unordered);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   195
    s = phase->transform(s);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   196
    mem->set_memory_at(fieldidx, s);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   197
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   198
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   199
  if (!finish_transform(phase, can_reshape, ctl, mem)) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   200
    return NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   201
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   202
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   203
  return mem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   204
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   205
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   206
bool ArrayCopyNode::prepare_array_copy(PhaseGVN *phase, bool can_reshape,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   207
                                       Node*& adr_src,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   208
                                       Node*& base_src,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   209
                                       Node*& adr_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   210
                                       Node*& base_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   211
                                       BasicType& copy_type,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   212
                                       const Type*& value_type,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   213
                                       bool& disjoint_bases) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   214
  Node* src = in(ArrayCopyNode::Src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   215
  Node* dest = in(ArrayCopyNode::Dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   216
  const Type* src_type = phase->type(src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   217
  const TypeAryPtr* ary_src = src_type->isa_aryptr();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   218
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   219
  if (is_arraycopy() || is_copyofrange() || is_copyof()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   220
    const Type* dest_type = phase->type(dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   221
    const TypeAryPtr* ary_dest = dest_type->isa_aryptr();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   222
    Node* src_offset = in(ArrayCopyNode::SrcPos);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   223
    Node* dest_offset = in(ArrayCopyNode::DestPos);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   224
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   225
    // newly allocated object is guaranteed to not overlap with source object
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   226
    disjoint_bases = is_alloc_tightly_coupled();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   227
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   228
    if (ary_src  == NULL || ary_src->klass()  == NULL ||
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   229
        ary_dest == NULL || ary_dest->klass() == NULL) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   230
      // We don't know if arguments are arrays
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   231
      return false;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   232
    }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   233
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   234
    BasicType src_elem  = ary_src->klass()->as_array_klass()->element_type()->basic_type();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   235
    BasicType dest_elem = ary_dest->klass()->as_array_klass()->element_type()->basic_type();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   236
    if (src_elem  == T_ARRAY)  src_elem  = T_OBJECT;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   237
    if (dest_elem == T_ARRAY)  dest_elem = T_OBJECT;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   238
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   239
    if (src_elem != dest_elem || dest_elem == T_VOID) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   240
      // We don't know if arguments are arrays of the same type
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   241
      return false;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   242
    }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   243
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   244
    if (dest_elem == T_OBJECT && (!is_alloc_tightly_coupled() || !GraphKit::use_ReduceInitialCardMarks())) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   245
      // It's an object array copy but we can't emit the card marking
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   246
      // that is needed
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   247
      return false;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   248
    }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   249
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   250
    value_type = ary_src->elem();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   251
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   252
    base_src = src;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   253
    base_dest = dest;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   254
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   255
    uint shift  = exact_log2(type2aelembytes(dest_elem));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   256
    uint header = arrayOopDesc::base_offset_in_bytes(dest_elem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   257
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   258
    adr_src = src;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   259
    adr_dest = dest;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   260
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   261
    src_offset = Compile::conv_I2X_index(phase, src_offset, ary_src->size());
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   262
    dest_offset = Compile::conv_I2X_index(phase, dest_offset, ary_dest->size());
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   263
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   264
    Node* src_scale = phase->transform(new LShiftXNode(src_offset, phase->intcon(shift)));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   265
    Node* dest_scale = phase->transform(new LShiftXNode(dest_offset, phase->intcon(shift)));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   266
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   267
    adr_src = phase->transform(new AddPNode(base_src, adr_src, src_scale));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   268
    adr_dest = phase->transform(new AddPNode(base_dest, adr_dest, dest_scale));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   269
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   270
    adr_src = new AddPNode(base_src, adr_src, phase->MakeConX(header));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   271
    adr_dest = new AddPNode(base_dest, adr_dest, phase->MakeConX(header));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   272
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   273
    adr_src = phase->transform(adr_src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   274
    adr_dest = phase->transform(adr_dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   275
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   276
    copy_type = dest_elem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   277
  } else {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   278
    assert (is_clonebasic(), "should be");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   279
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   280
    disjoint_bases = true;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   281
    assert(src->is_AddP(), "should be base + off");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   282
    assert(dest->is_AddP(), "should be base + off");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   283
    adr_src = src;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   284
    base_src = src->in(AddPNode::Base);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   285
    adr_dest = dest;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   286
    base_dest = dest->in(AddPNode::Base);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   287
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   288
    assert(phase->type(src->in(AddPNode::Offset))->is_intptr_t()->get_con() == phase->type(dest->in(AddPNode::Offset))->is_intptr_t()->get_con(), "same start offset?");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   289
    BasicType elem = ary_src->klass()->as_array_klass()->element_type()->basic_type();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   290
    if (elem == T_ARRAY)  elem = T_OBJECT;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   291
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   292
    int diff = arrayOopDesc::base_offset_in_bytes(elem) - phase->type(src->in(AddPNode::Offset))->is_intptr_t()->get_con();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   293
    assert(diff >= 0, "clone should not start after 1st array element");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   294
    if (diff > 0) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   295
      adr_src = phase->transform(new AddPNode(base_src, adr_src, phase->MakeConX(diff)));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   296
      adr_dest = phase->transform(new AddPNode(base_dest, adr_dest, phase->MakeConX(diff)));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   297
    }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   298
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   299
    copy_type = elem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   300
    value_type = ary_src->elem();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   301
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   302
  return true;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   303
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   304
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   305
const TypePtr* ArrayCopyNode::get_address_type(PhaseGVN *phase, Node* n) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   306
  const Type* at = phase->type(n);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   307
  assert(at != Type::TOP, "unexpected type");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   308
  const TypePtr* atp = at->isa_ptr();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   309
  // adjust atp to be the correct array element address type
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   310
  atp = atp->add_offset(Type::OffsetBot);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   311
  return atp;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   312
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   313
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   314
void ArrayCopyNode::array_copy_test_overlap(PhaseGVN *phase, bool can_reshape, bool disjoint_bases, int count, Node*& forward_ctl, Node*& backward_ctl) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   315
  Node* ctl = in(TypeFunc::Control);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   316
  if (!disjoint_bases && count > 1) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   317
    Node* src_offset = in(ArrayCopyNode::SrcPos);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   318
    Node* dest_offset = in(ArrayCopyNode::DestPos);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   319
    assert(src_offset != NULL && dest_offset != NULL, "should be");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   320
    Node* cmp = phase->transform(new CmpINode(src_offset, dest_offset));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   321
    Node *bol = phase->transform(new BoolNode(cmp, BoolTest::lt));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   322
    IfNode *iff = new IfNode(ctl, bol, PROB_FAIR, COUNT_UNKNOWN);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   323
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   324
    phase->transform(iff);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   325
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   326
    forward_ctl = phase->transform(new IfFalseNode(iff));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   327
    backward_ctl = phase->transform(new IfTrueNode(iff));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   328
  } else {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   329
    forward_ctl = ctl;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   330
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   331
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   332
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   333
Node* ArrayCopyNode::array_copy_forward(PhaseGVN *phase,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   334
                                        bool can_reshape,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   335
                                        Node* forward_ctl,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   336
                                        Node* start_mem_src,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   337
                                        Node* start_mem_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   338
                                        const TypePtr* atp_src,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   339
                                        const TypePtr* atp_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   340
                                        Node* adr_src,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   341
                                        Node* base_src,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   342
                                        Node* adr_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   343
                                        Node* base_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   344
                                        BasicType copy_type,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   345
                                        const Type* value_type,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   346
                                        int count) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   347
  Node* mem = phase->C->top();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   348
  if (!forward_ctl->is_top()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   349
    // copy forward
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   350
    mem = start_mem_dest;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   351
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   352
    if (count > 0) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   353
      Node* v = LoadNode::make(*phase, forward_ctl, start_mem_src, adr_src, atp_src, value_type, copy_type, MemNode::unordered);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   354
      v = phase->transform(v);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   355
      mem = StoreNode::make(*phase, forward_ctl, mem, adr_dest, atp_dest, v, copy_type, MemNode::unordered);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   356
      mem = phase->transform(mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   357
      for (int i = 1; i < count; i++) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   358
        Node* off  = phase->MakeConX(type2aelembytes(copy_type) * i);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   359
        Node* next_src = phase->transform(new AddPNode(base_src,adr_src,off));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   360
        Node* next_dest = phase->transform(new AddPNode(base_dest,adr_dest,off));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   361
        v = LoadNode::make(*phase, forward_ctl, mem, next_src, atp_src, value_type, copy_type, MemNode::unordered);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   362
        v = phase->transform(v);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   363
        mem = StoreNode::make(*phase, forward_ctl,mem,next_dest,atp_dest,v, copy_type, MemNode::unordered);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   364
        mem = phase->transform(mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   365
      }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   366
    } else if(can_reshape) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   367
      PhaseIterGVN* igvn = phase->is_IterGVN();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   368
      igvn->_worklist.push(adr_src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   369
      igvn->_worklist.push(adr_dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   370
    }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   371
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   372
  return mem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   373
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   374
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   375
Node* ArrayCopyNode::array_copy_backward(PhaseGVN *phase,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   376
                                         bool can_reshape,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   377
                                         Node* backward_ctl,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   378
                                         Node* start_mem_src,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   379
                                         Node* start_mem_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   380
                                         const TypePtr* atp_src,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   381
                                         const TypePtr* atp_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   382
                                         Node* adr_src,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   383
                                         Node* base_src,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   384
                                         Node* adr_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   385
                                         Node* base_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   386
                                         BasicType copy_type,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   387
                                         const Type* value_type,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   388
                                         int count) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   389
  Node* mem = phase->C->top();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   390
  if (!backward_ctl->is_top()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   391
    // copy backward
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   392
    mem = start_mem_dest;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   393
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   394
    if (count > 0) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   395
      for (int i = count-1; i >= 1; i--) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   396
        Node* off  = phase->MakeConX(type2aelembytes(copy_type) * i);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   397
        Node* next_src = phase->transform(new AddPNode(base_src,adr_src,off));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   398
        Node* next_dest = phase->transform(new AddPNode(base_dest,adr_dest,off));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   399
        Node* v = LoadNode::make(*phase, backward_ctl, mem, next_src, atp_src, value_type, copy_type, MemNode::unordered);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   400
        v = phase->transform(v);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   401
        mem = StoreNode::make(*phase, backward_ctl,mem,next_dest,atp_dest,v, copy_type, MemNode::unordered);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   402
        mem = phase->transform(mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   403
      }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   404
      Node* v = LoadNode::make(*phase, backward_ctl, mem, adr_src, atp_src, value_type, copy_type, MemNode::unordered);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   405
      v = phase->transform(v);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   406
      mem = StoreNode::make(*phase, backward_ctl, mem, adr_dest, atp_dest, v, copy_type, MemNode::unordered);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   407
      mem = phase->transform(mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   408
    } else if(can_reshape) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   409
      PhaseIterGVN* igvn = phase->is_IterGVN();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   410
      igvn->_worklist.push(adr_src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   411
      igvn->_worklist.push(adr_dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   412
    }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   413
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   414
  return mem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   415
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   416
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   417
bool ArrayCopyNode::finish_transform(PhaseGVN *phase, bool can_reshape,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   418
                                     Node* ctl, Node *mem) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   419
  if (can_reshape) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   420
    PhaseIterGVN* igvn = phase->is_IterGVN();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   421
    igvn->set_delay_transform(false);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   422
    if (is_clonebasic()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   423
      Node* out_mem = proj_out(TypeFunc::Memory);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   424
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   425
      if (out_mem->outcnt() != 1 || !out_mem->raw_out(0)->is_MergeMem() ||
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   426
          out_mem->raw_out(0)->outcnt() != 1 || !out_mem->raw_out(0)->raw_out(0)->is_MemBar()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   427
        assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   428
        return false;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   429
      }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   430
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   431
      igvn->replace_node(out_mem->raw_out(0), mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   432
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   433
      Node* out_ctl = proj_out(TypeFunc::Control);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   434
      igvn->replace_node(out_ctl, ctl);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   435
    } else {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   436
      // replace fallthrough projections of the ArrayCopyNode by the
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   437
      // new memory, control and the input IO.
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   438
      CallProjections callprojs;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   439
      extract_projections(&callprojs, true);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   440
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   441
      igvn->replace_node(callprojs.fallthrough_ioproj, in(TypeFunc::I_O));
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   442
      igvn->replace_node(callprojs.fallthrough_memproj, mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   443
      igvn->replace_node(callprojs.fallthrough_catchproj, ctl);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   444
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   445
      // The ArrayCopyNode is not disconnected. It still has the
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   446
      // projections for the exception case. Replace current
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   447
      // ArrayCopyNode with a dummy new one with a top() control so
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   448
      // that this part of the graph stays consistent but is
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   449
      // eventually removed.
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   450
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   451
      set_req(0, phase->C->top());
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   452
      remove_dead_region(phase, can_reshape);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   453
    }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   454
  } else {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   455
    if (in(TypeFunc::Control) != ctl) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   456
      // we can't return new memory and control from Ideal at parse time
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   457
      assert(!is_clonebasic(), "added control for clone?");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   458
      return NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   459
    }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   460
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   461
  return true;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   462
}
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   463
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   464
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   465
Node *ArrayCopyNode::Ideal(PhaseGVN *phase, bool can_reshape) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   466
  if (remove_dead_region(phase, can_reshape))  return this;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   467
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   468
  if (StressArrayCopyMacroNode && !can_reshape) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   469
    phase->record_for_igvn(this);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   470
    return NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   471
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   472
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   473
  // See if it's a small array copy and we can inline it as
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   474
  // loads/stores
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   475
  // Here we can only do:
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   476
  // - arraycopy if all arguments were validated before and we don't
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   477
  // need card marking
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   478
  // - clone for which we don't need to do card marking
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   479
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   480
  if (!is_clonebasic() && !is_arraycopy_validated() &&
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   481
      !is_copyofrange_validated() && !is_copyof_validated()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   482
    return NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   483
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   484
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   485
  if (in(TypeFunc::Control)->is_top() || in(TypeFunc::Memory)->is_top()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   486
    return NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   487
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   488
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   489
  int count = get_count(phase);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   490
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   491
  if (count < 0 || count > ArrayCopyLoadStoreMaxElem) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   492
    return NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   493
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   494
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   495
  Node* mem = try_clone_instance(phase, can_reshape, count);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   496
  if (mem != NULL) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   497
    return mem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   498
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   499
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   500
  Node* adr_src = NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   501
  Node* base_src = NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   502
  Node* adr_dest = NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   503
  Node* base_dest = NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   504
  BasicType copy_type = T_ILLEGAL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   505
  const Type* value_type = NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   506
  bool disjoint_bases = false;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   507
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   508
  if (!prepare_array_copy(phase, can_reshape,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   509
                          adr_src, base_src, adr_dest, base_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   510
                          copy_type, value_type, disjoint_bases)) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   511
    return NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   512
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   513
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   514
  Node* src = in(ArrayCopyNode::Src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   515
  Node* dest = in(ArrayCopyNode::Dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   516
  const TypePtr* atp_src = get_address_type(phase, src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   517
  const TypePtr* atp_dest = get_address_type(phase, dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   518
  uint alias_idx_src = phase->C->get_alias_index(atp_src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   519
  uint alias_idx_dest = phase->C->get_alias_index(atp_dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   520
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   521
  Node *in_mem = in(TypeFunc::Memory);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   522
  Node *start_mem_src = in_mem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   523
  Node *start_mem_dest = in_mem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   524
  if (in_mem->is_MergeMem()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   525
    start_mem_src = in_mem->as_MergeMem()->memory_at(alias_idx_src);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   526
    start_mem_dest = in_mem->as_MergeMem()->memory_at(alias_idx_dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   527
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   528
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   529
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   530
  if (can_reshape) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   531
    assert(!phase->is_IterGVN()->delay_transform(), "cannot delay transforms");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   532
    phase->is_IterGVN()->set_delay_transform(true);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   533
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   534
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   535
  Node* backward_ctl = phase->C->top();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   536
  Node* forward_ctl = phase->C->top();
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   537
  array_copy_test_overlap(phase, can_reshape, disjoint_bases, count, forward_ctl, backward_ctl);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   538
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   539
  Node* forward_mem = array_copy_forward(phase, can_reshape, forward_ctl,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   540
                                         start_mem_src, start_mem_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   541
                                         atp_src, atp_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   542
                                         adr_src, base_src, adr_dest, base_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   543
                                         copy_type, value_type, count);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   544
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   545
  Node* backward_mem = array_copy_backward(phase, can_reshape, backward_ctl,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   546
                                           start_mem_src, start_mem_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   547
                                           atp_src, atp_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   548
                                           adr_src, base_src, adr_dest, base_dest,
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   549
                                           copy_type, value_type, count);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   550
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   551
  Node* ctl = NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   552
  if (!forward_ctl->is_top() && !backward_ctl->is_top()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   553
    ctl = new RegionNode(3);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   554
    mem = new PhiNode(ctl, Type::MEMORY, atp_dest);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   555
    ctl->init_req(1, forward_ctl);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   556
    mem->init_req(1, forward_mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   557
    ctl->init_req(2, backward_ctl);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   558
    mem->init_req(2, backward_mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   559
    ctl = phase->transform(ctl);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   560
    mem = phase->transform(mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   561
  } else if (!forward_ctl->is_top()) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   562
    ctl = forward_ctl;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   563
    mem = forward_mem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   564
  } else {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   565
    assert(!backward_ctl->is_top(), "no copy?");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   566
    ctl = backward_ctl;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   567
    mem = backward_mem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   568
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   569
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   570
  if (can_reshape) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   571
    assert(phase->is_IterGVN()->delay_transform(), "should be delaying transforms");
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   572
    phase->is_IterGVN()->set_delay_transform(false);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   573
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   574
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   575
  MergeMemNode* out_mem = MergeMemNode::make(in_mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   576
  out_mem->set_memory_at(alias_idx_dest, mem);
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   577
  mem = out_mem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   578
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   579
  if (!finish_transform(phase, can_reshape, ctl, mem)) {
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   580
    return NULL;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   581
  }
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   582
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   583
  return mem;
ef2be52deeaf 6912521: System.arraycopy works slower than the simple loop for little lengths
roland
parents:
diff changeset
   584
}