1 /*
   2  * Copyright (c) 2012, 2016, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "opto/arraycopynode.hpp"
  27 #include "oops/objArrayKlass.hpp"
  28 #include "opto/convertnode.hpp"
  29 #include "opto/graphKit.hpp"
  30 #include "opto/macro.hpp"
  31 #include "opto/runtime.hpp"
  32 
  33 
  34 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, Node* precedent) {
  35   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
  36   mb->init_req(TypeFunc::Control, *ctrl);
  37   mb->init_req(TypeFunc::Memory, *mem);
  38   transform_later(mb);
  39   *ctrl = new ProjNode(mb,TypeFunc::Control);
  40   transform_later(*ctrl);
  41   Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
  42   transform_later(mem_proj);
  43   *mem = mem_proj;
  44 }
  45 
  46 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
  47   uint shift  = exact_log2(type2aelembytes(elembt));
  48   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
  49   Node* base =  basic_plus_adr(ary, header);
  50 #ifdef _LP64
  51   // see comment in GraphKit::array_element_address
  52   int index_max = max_jint - 1;  // array size is max_jint, index is one less
  53   const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
  54   idx = transform_later( new ConvI2LNode(idx, lidxtype) );
  55 #endif
  56   Node* scale = new LShiftXNode(idx, intcon(shift));
  57   transform_later(scale);
  58   return basic_plus_adr(ary, base, scale);
  59 }
  60 
  61 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
  62   return transform_later(new ConvI2LNode(offset));
  63 }
  64 
  65 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
  66                                        const TypeFunc* call_type, address call_addr,
  67                                        const char* call_name,
  68                                        const TypePtr* adr_type,
  69                                        Node* parm0, Node* parm1,
  70                                        Node* parm2, Node* parm3,
  71                                        Node* parm4, Node* parm5,
  72                                        Node* parm6, Node* parm7) {
  73   int size = call_type->domain()->cnt();
  74   Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
  75   call->init_req(TypeFunc::Control, ctrl);
  76   call->init_req(TypeFunc::I_O    , top());
  77   call->init_req(TypeFunc::Memory , mem);
  78   call->init_req(TypeFunc::ReturnAdr, top());
  79   call->init_req(TypeFunc::FramePtr, top());
  80 
  81   // Hook each parm in order.  Stop looking at the first NULL.
  82   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
  83   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
  84   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
  85   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
  86   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
  87   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
  88   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
  89   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
  90     /* close each nested if ===> */  } } } } } } } }
  91   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
  92 
  93   return call;
  94 }
  95 
  96 
  97 //------------------------------generate_guard---------------------------
  98 // Helper function for generating guarded fast-slow graph structures.
  99 // The given 'test', if true, guards a slow path.  If the test fails
 100 // then a fast path can be taken.  (We generally hope it fails.)
 101 // In all cases, GraphKit::control() is updated to the fast path.
 102 // The returned value represents the control for the slow path.
 103 // The return value is never 'top'; it is either a valid control
 104 // or NULL if it is obvious that the slow path can never be taken.
 105 // Also, if region and the slow control are not NULL, the slow edge
 106 // is appended to the region.
 107 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
 108   if ((*ctrl)->is_top()) {
 109     // Already short circuited.
 110     return NULL;
 111   }
 112   // Build an if node and its projections.
 113   // If test is true we take the slow path, which we assume is uncommon.
 114   if (_igvn.type(test) == TypeInt::ZERO) {
 115     // The slow branch is never taken.  No need to build this guard.
 116     return NULL;
 117   }
 118 
 119   IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
 120   transform_later(iff);
 121 
 122   Node* if_slow = new IfTrueNode(iff);
 123   transform_later(if_slow);
 124 
 125   if (region != NULL) {
 126     region->add_req(if_slow);
 127   }
 128 
 129   Node* if_fast = new IfFalseNode(iff);
 130   transform_later(if_fast);
 131 
 132   *ctrl = if_fast;
 133 
 134   return if_slow;
 135 }
 136 
 137 inline Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
 138   return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
 139 }
 140 
 141 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
 142   if ((*ctrl)->is_top())
 143     return;                // already stopped
 144   if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
 145     return;                // index is already adequately typed
 146   Node* cmp_lt = new CmpINode(index, intcon(0));
 147   transform_later(cmp_lt);
 148   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 149   transform_later(bol_lt);
 150   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 151 }
 152 
 153 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
 154   if ((*ctrl)->is_top())
 155     return;                // already stopped
 156   bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
 157   if (zero_offset && subseq_length->eqv_uncast(array_length))
 158     return;                // common case of whole-array copy
 159   Node* last = subseq_length;
 160   if (!zero_offset) {            // last += offset
 161     last = new AddINode(last, offset);
 162     transform_later(last);
 163   }
 164   Node* cmp_lt = new CmpUNode(array_length, last);
 165   transform_later(cmp_lt);
 166   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 167   transform_later(bol_lt);
 168   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 169 }
 170 
 171 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
 172   if ((*ctrl)->is_top())  return NULL;
 173 
 174   if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
 175     return NULL;                // index is already adequately typed
 176   Node* cmp_le = new CmpINode(index, intcon(0));
 177   transform_later(cmp_le);
 178   BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
 179   Node* bol_le = new BoolNode(cmp_le, le_or_eq);
 180   transform_later(bol_le);
 181   Node* is_notp = generate_guard(ctrl, bol_le, NULL, PROB_MIN);
 182 
 183   return is_notp;
 184 }
 185 
 186 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
 187   transform_later(call);
 188 
 189   *ctrl = new ProjNode(call,TypeFunc::Control);
 190   transform_later(*ctrl);
 191   Node* newmem = new ProjNode(call, TypeFunc::Memory);
 192   transform_later(newmem);
 193 
 194   uint alias_idx = C->get_alias_index(adr_type);
 195   if (alias_idx != Compile::AliasIdxBot) {
 196     *mem = MergeMemNode::make(*mem);
 197     (*mem)->set_memory_at(alias_idx, newmem);
 198   } else {
 199     *mem = MergeMemNode::make(newmem);
 200   }
 201   transform_later(*mem);
 202 }
 203 
 204 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
 205                                               Node* src_offset,
 206                                               Node* dest_offset,
 207                                               bool disjoint_bases,
 208                                               const char* &name,
 209                                               bool dest_uninitialized) {
 210   const TypeInt* src_offset_inttype  = _igvn.find_int_type(src_offset);;
 211   const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);;
 212 
 213   bool aligned = false;
 214   bool disjoint = disjoint_bases;
 215 
 216   // if the offsets are the same, we can treat the memory regions as
 217   // disjoint, because either the memory regions are in different arrays,
 218   // or they are identical (which we can treat as disjoint.)  We can also
 219   // treat a copy with a destination index  less that the source index
 220   // as disjoint since a low->high copy will work correctly in this case.
 221   if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
 222       dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
 223     // both indices are constants
 224     int s_offs = src_offset_inttype->get_con();
 225     int d_offs = dest_offset_inttype->get_con();
 226     int element_size = type2aelembytes(t);
 227     aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
 228               ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
 229     if (s_offs >= d_offs)  disjoint = true;
 230   } else if (src_offset == dest_offset && src_offset != NULL) {
 231     // This can occur if the offsets are identical non-constants.
 232     disjoint = true;
 233   }
 234 
 235   return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
 236 }
 237 
 238 #define XTOP LP64_ONLY(COMMA top())
 239 
 240 // Generate an optimized call to arraycopy.
 241 // Caller must guard against non-arrays.
 242 // Caller must determine a common array basic-type for both arrays.
 243 // Caller must validate offsets against array bounds.
 244 // The slow_region has already collected guard failure paths
 245 // (such as out of bounds length or non-conformable array types).
 246 // The generated code has this shape, in general:
 247 //
 248 //     if (length == 0)  return   // via zero_path
 249 //     slowval = -1
 250 //     if (types unknown) {
 251 //       slowval = call generic copy loop
 252 //       if (slowval == 0)  return  // via checked_path
 253 //     } else if (indexes in bounds) {
 254 //       if ((is object array) && !(array type check)) {
 255 //         slowval = call checked copy loop
 256 //         if (slowval == 0)  return  // via checked_path
 257 //       } else {
 258 //         call bulk copy loop
 259 //         return  // via fast_path
 260 //       }
 261 //     }
 262 //     // adjust params for remaining work:
 263 //     if (slowval != -1) {
 264 //       n = -1^slowval; src_offset += n; dest_offset += n; length -= n
 265 //     }
 266 //   slow_region:
 267 //     call slow arraycopy(src, src_offset, dest, dest_offset, length)
 268 //     return  // via slow_call_path
 269 //
 270 // This routine is used from several intrinsics:  System.arraycopy,
 271 // Object.clone (the array subcase), and Arrays.copyOf[Range].
 272 //
 273 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
 274                                            Node** ctrl, MergeMemNode* mem, Node** io,
 275                                            const TypePtr* adr_type,
 276                                            BasicType basic_elem_type,
 277                                            Node* src,  Node* src_offset,
 278                                            Node* dest, Node* dest_offset,
 279                                            Node* copy_length,
 280                                            bool disjoint_bases,
 281                                            bool length_never_negative,
 282                                            RegionNode* slow_region) {
 283   if (slow_region == NULL) {
 284     slow_region = new RegionNode(1);
 285     transform_later(slow_region);
 286   }
 287 
 288   Node* original_dest      = dest;
 289   bool  dest_uninitialized = false;
 290 
 291   // See if this is the initialization of a newly-allocated array.
 292   // If so, we will take responsibility here for initializing it to zero.
 293   // (Note:  Because tightly_coupled_allocation performs checks on the
 294   // out-edges of the dest, we need to avoid making derived pointers
 295   // from it until we have checked its uses.)
 296   if (ReduceBulkZeroing
 297       && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
 298       && basic_elem_type != T_CONFLICT // avoid corner case
 299       && !src->eqv_uncast(dest)
 300       && alloc != NULL
 301       && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
 302       && alloc->maybe_set_complete(&_igvn)) {
 303     // "You break it, you buy it."
 304     InitializeNode* init = alloc->initialization();
 305     assert(init->is_complete(), "we just did this");
 306     init->set_complete_with_arraycopy();
 307     assert(dest->is_CheckCastPP(), "sanity");
 308     assert(dest->in(0)->in(0) == init, "dest pinned");
 309     adr_type = TypeRawPtr::BOTTOM;  // all initializations are into raw memory
 310     // From this point on, every exit path is responsible for
 311     // initializing any non-copied parts of the object to zero.
 312     // Also, if this flag is set we make sure that arraycopy interacts properly
 313     // with G1, eliding pre-barriers. See CR 6627983.
 314     dest_uninitialized = true;
 315   } else {
 316     // No zeroing elimination here.
 317     alloc             = NULL;
 318     //original_dest   = dest;
 319     //dest_uninitialized = false;
 320   }
 321 
 322   uint alias_idx = C->get_alias_index(adr_type);
 323 
 324   // Results are placed here:
 325   enum { fast_path        = 1,  // normal void-returning assembly stub
 326          checked_path     = 2,  // special assembly stub with cleanup
 327          slow_call_path   = 3,  // something went wrong; call the VM
 328          zero_path        = 4,  // bypass when length of copy is zero
 329          bcopy_path       = 5,  // copy primitive array by 64-bit blocks
 330          PATH_LIMIT       = 6
 331   };
 332   RegionNode* result_region = new RegionNode(PATH_LIMIT);
 333   PhiNode*    result_i_o    = new PhiNode(result_region, Type::ABIO);
 334   PhiNode*    result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
 335   assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
 336   transform_later(result_region);
 337   transform_later(result_i_o);
 338   transform_later(result_memory);
 339 
 340   // The slow_control path:
 341   Node* slow_control;
 342   Node* slow_i_o = *io;
 343   Node* slow_mem = mem->memory_at(alias_idx);
 344   DEBUG_ONLY(slow_control = (Node*) badAddress);
 345 
 346   // Checked control path:
 347   Node* checked_control = top();
 348   Node* checked_mem     = NULL;
 349   Node* checked_i_o     = NULL;
 350   Node* checked_value   = NULL;
 351 
 352   if (basic_elem_type == T_CONFLICT) {
 353     assert(!dest_uninitialized, "");
 354     Node* cv = generate_generic_arraycopy(ctrl, &mem,
 355                                           adr_type,
 356                                           src, src_offset, dest, dest_offset,
 357                                           copy_length, dest_uninitialized);
 358     if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
 359     checked_control = *ctrl;
 360     checked_i_o     = *io;
 361     checked_mem     = mem->memory_at(alias_idx);
 362     checked_value   = cv;
 363     *ctrl = top();
 364   }
 365 
 366   Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
 367   if (not_pos != NULL) {
 368     Node* local_ctrl = not_pos, *local_io = *io;
 369     MergeMemNode* local_mem = MergeMemNode::make(mem);
 370     transform_later(local_mem);
 371 
 372     // (6) length must not be negative.
 373     if (!length_never_negative) {
 374       generate_negative_guard(&local_ctrl, copy_length, slow_region);
 375     }
 376 
 377     // copy_length is 0.
 378     if (dest_uninitialized) {
 379       assert(!local_ctrl->is_top(), "no ctrl?");
 380       Node* dest_length = alloc->in(AllocateNode::ALength);
 381       if (copy_length->eqv_uncast(dest_length)
 382           || _igvn.find_int_con(dest_length, 1) <= 0) {
 383         // There is no zeroing to do. No need for a secondary raw memory barrier.
 384       } else {
 385         // Clear the whole thing since there are no source elements to copy.
 386         generate_clear_array(local_ctrl, local_mem,
 387                              adr_type, dest, basic_elem_type,
 388                              intcon(0), NULL,
 389                              alloc->in(AllocateNode::AllocSize));
 390         // Use a secondary InitializeNode as raw memory barrier.
 391         // Currently it is needed only on this path since other
 392         // paths have stub or runtime calls as raw memory barriers.
 393         MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
 394                                           Compile::AliasIdxRaw,
 395                                           top());
 396         transform_later(mb);
 397         mb->set_req(TypeFunc::Control,local_ctrl);
 398         mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
 399         local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
 400         local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
 401 
 402         InitializeNode* init = mb->as_Initialize();
 403         init->set_complete(&_igvn);  // (there is no corresponding AllocateNode)
 404       }
 405     }
 406 
 407     // Present the results of the fast call.
 408     result_region->init_req(zero_path, local_ctrl);
 409     result_i_o   ->init_req(zero_path, local_io);
 410     result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
 411   }
 412 
 413   if (!(*ctrl)->is_top() && dest_uninitialized) {
 414     // We have to initialize the *uncopied* part of the array to zero.
 415     // The copy destination is the slice dest[off..off+len].  The other slices
 416     // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
 417     Node* dest_size   = alloc->in(AllocateNode::AllocSize);
 418     Node* dest_length = alloc->in(AllocateNode::ALength);
 419     Node* dest_tail   = transform_later( new AddINode(dest_offset, copy_length));
 420 
 421     // If there is a head section that needs zeroing, do it now.
 422     if (_igvn.find_int_con(dest_offset, -1) != 0) {
 423       generate_clear_array(*ctrl, mem,
 424                            adr_type, dest, basic_elem_type,
 425                            intcon(0), dest_offset,
 426                            NULL);
 427     }
 428 
 429     // Next, perform a dynamic check on the tail length.
 430     // It is often zero, and we can win big if we prove this.
 431     // There are two wins:  Avoid generating the ClearArray
 432     // with its attendant messy index arithmetic, and upgrade
 433     // the copy to a more hardware-friendly word size of 64 bits.
 434     Node* tail_ctl = NULL;
 435     if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
 436       Node* cmp_lt   = transform_later( new CmpINode(dest_tail, dest_length) );
 437       Node* bol_lt   = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
 438       tail_ctl = generate_slow_guard(ctrl, bol_lt, NULL);
 439       assert(tail_ctl != NULL || !(*ctrl)->is_top(), "must be an outcome");
 440     }
 441 
 442     // At this point, let's assume there is no tail.
 443     if (!(*ctrl)->is_top() && alloc != NULL && basic_elem_type != T_OBJECT) {
 444       // There is no tail.  Try an upgrade to a 64-bit copy.
 445       bool didit = false;
 446       {
 447         Node* local_ctrl = *ctrl, *local_io = *io;
 448         MergeMemNode* local_mem = MergeMemNode::make(mem);
 449         transform_later(local_mem);
 450 
 451         didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
 452                                          adr_type, basic_elem_type, alloc,
 453                                          src, src_offset, dest, dest_offset,
 454                                          dest_size, dest_uninitialized);
 455         if (didit) {
 456           // Present the results of the block-copying fast call.
 457           result_region->init_req(bcopy_path, local_ctrl);
 458           result_i_o   ->init_req(bcopy_path, local_io);
 459           result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
 460         }
 461       }
 462       if (didit) {
 463         *ctrl = top();     // no regular fast path
 464       }
 465     }
 466 
 467     // Clear the tail, if any.
 468     if (tail_ctl != NULL) {
 469       Node* notail_ctl = (*ctrl)->is_top() ? NULL : *ctrl;
 470       *ctrl = tail_ctl;
 471       if (notail_ctl == NULL) {
 472         generate_clear_array(*ctrl, mem,
 473                              adr_type, dest, basic_elem_type,
 474                              dest_tail, NULL,
 475                              dest_size);
 476       } else {
 477         // Make a local merge.
 478         Node* done_ctl = transform_later(new RegionNode(3));
 479         Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
 480         done_ctl->init_req(1, notail_ctl);
 481         done_mem->init_req(1, mem->memory_at(alias_idx));
 482         generate_clear_array(*ctrl, mem,
 483                              adr_type, dest, basic_elem_type,
 484                              dest_tail, NULL,
 485                              dest_size);
 486         done_ctl->init_req(2, *ctrl);
 487         done_mem->init_req(2, mem->memory_at(alias_idx));
 488         *ctrl = done_ctl;
 489         mem->set_memory_at(alias_idx, done_mem);
 490       }
 491     }
 492   }
 493 
 494   BasicType copy_type = basic_elem_type;
 495   assert(basic_elem_type != T_ARRAY, "caller must fix this");
 496   if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
 497     // If src and dest have compatible element types, we can copy bits.
 498     // Types S[] and D[] are compatible if D is a supertype of S.
 499     //
 500     // If they are not, we will use checked_oop_disjoint_arraycopy,
 501     // which performs a fast optimistic per-oop check, and backs off
 502     // further to JVM_ArrayCopy on the first per-oop check that fails.
 503     // (Actually, we don't move raw bits only; the GC requires card marks.)
 504 
 505     // We don't need a subtype check for validated copies and Object[].clone()
 506     bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
 507                               ac->is_copyofrange_validated() || ac->is_cloneoop();
 508     if (!skip_subtype_check) {
 509       // Get the klass* for both src and dest
 510       Node* src_klass  = ac->in(ArrayCopyNode::SrcKlass);
 511       Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
 512 
 513       assert(src_klass != NULL && dest_klass != NULL, "should have klasses");
 514 
 515       // Generate the subtype check.
 516       // This might fold up statically, or then again it might not.
 517       //
 518       // Non-static example:  Copying List<String>.elements to a new String[].
 519       // The backing store for a List<String> is always an Object[],
 520       // but its elements are always type String, if the generic types
 521       // are correct at the source level.
 522       //
 523       // Test S[] against D[], not S against D, because (probably)
 524       // the secondary supertype cache is less busy for S[] than S.
 525       // This usually only matters when D is an interface.
 526       Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, &_igvn);
 527       // Plug failing path into checked_oop_disjoint_arraycopy
 528       if (not_subtype_ctrl != top()) {
 529         Node* local_ctrl = not_subtype_ctrl;
 530         MergeMemNode* local_mem = MergeMemNode::make(mem);
 531         transform_later(local_mem);
 532 
 533         // (At this point we can assume disjoint_bases, since types differ.)
 534         int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
 535         Node* p1 = basic_plus_adr(dest_klass, ek_offset);
 536         Node* n1 = LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
 537         Node* dest_elem_klass = transform_later(n1);
 538         Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
 539                                                 adr_type,
 540                                                 dest_elem_klass,
 541                                                 src, src_offset, dest, dest_offset,
 542                                                 ConvI2X(copy_length), dest_uninitialized);
 543         if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
 544         checked_control = local_ctrl;
 545         checked_i_o     = *io;
 546         checked_mem     = local_mem->memory_at(alias_idx);
 547         checked_value   = cv;
 548       }
 549     }
 550     // At this point we know we do not need type checks on oop stores.
 551 
 552     // Let's see if we need card marks:
 553     if (alloc != NULL && GraphKit::use_ReduceInitialCardMarks()) {
 554       // If we do not need card marks, copy using the jint or jlong stub.
 555       copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
 556       assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
 557              "sizes agree");
 558     }
 559   }
 560 
 561   if (!(*ctrl)->is_top()) {
 562     // Generate the fast path, if possible.
 563     Node* local_ctrl = *ctrl;
 564     MergeMemNode* local_mem = MergeMemNode::make(mem);
 565     transform_later(local_mem);
 566 
 567     generate_unchecked_arraycopy(&local_ctrl, &local_mem,
 568                                  adr_type, copy_type, disjoint_bases,
 569                                  src, src_offset, dest, dest_offset,
 570                                  ConvI2X(copy_length), dest_uninitialized);
 571 
 572     // Present the results of the fast call.
 573     result_region->init_req(fast_path, local_ctrl);
 574     result_i_o   ->init_req(fast_path, *io);
 575     result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
 576   }
 577 
 578   // Here are all the slow paths up to this point, in one bundle:
 579   assert(slow_region != NULL, "allocated on entry");
 580   slow_control = slow_region;
 581   DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
 582 
 583   *ctrl = checked_control;
 584   if (!(*ctrl)->is_top()) {
 585     // Clean up after the checked call.
 586     // The returned value is either 0 or -1^K,
 587     // where K = number of partially transferred array elements.
 588     Node* cmp = new CmpINode(checked_value, intcon(0));
 589     transform_later(cmp);
 590     Node* bol = new BoolNode(cmp, BoolTest::eq);
 591     transform_later(bol);
 592     IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
 593     transform_later(iff);
 594 
 595     // If it is 0, we are done, so transfer to the end.
 596     Node* checks_done = new IfTrueNode(iff);
 597     transform_later(checks_done);
 598     result_region->init_req(checked_path, checks_done);
 599     result_i_o   ->init_req(checked_path, checked_i_o);
 600     result_memory->init_req(checked_path, checked_mem);
 601 
 602     // If it is not zero, merge into the slow call.
 603     *ctrl = new IfFalseNode(iff);
 604     transform_later(*ctrl);
 605     RegionNode* slow_reg2 = new RegionNode(3);
 606     PhiNode*    slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
 607     PhiNode*    slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
 608     transform_later(slow_reg2);
 609     transform_later(slow_i_o2);
 610     transform_later(slow_mem2);
 611     slow_reg2  ->init_req(1, slow_control);
 612     slow_i_o2  ->init_req(1, slow_i_o);
 613     slow_mem2  ->init_req(1, slow_mem);
 614     slow_reg2  ->init_req(2, *ctrl);
 615     slow_i_o2  ->init_req(2, checked_i_o);
 616     slow_mem2  ->init_req(2, checked_mem);
 617 
 618     slow_control = slow_reg2;
 619     slow_i_o     = slow_i_o2;
 620     slow_mem     = slow_mem2;
 621 
 622     if (alloc != NULL) {
 623       // We'll restart from the very beginning, after zeroing the whole thing.
 624       // This can cause double writes, but that's OK since dest is brand new.
 625       // So we ignore the low 31 bits of the value returned from the stub.
 626     } else {
 627       // We must continue the copy exactly where it failed, or else
 628       // another thread might see the wrong number of writes to dest.
 629       Node* checked_offset = new XorINode(checked_value, intcon(-1));
 630       Node* slow_offset    = new PhiNode(slow_reg2, TypeInt::INT);
 631       transform_later(checked_offset);
 632       transform_later(slow_offset);
 633       slow_offset->init_req(1, intcon(0));
 634       slow_offset->init_req(2, checked_offset);
 635 
 636       // Adjust the arguments by the conditionally incoming offset.
 637       Node* src_off_plus  = new AddINode(src_offset,  slow_offset);
 638       transform_later(src_off_plus);
 639       Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
 640       transform_later(dest_off_plus);
 641       Node* length_minus  = new SubINode(copy_length, slow_offset);
 642       transform_later(length_minus);
 643 
 644       // Tweak the node variables to adjust the code produced below:
 645       src_offset  = src_off_plus;
 646       dest_offset = dest_off_plus;
 647       copy_length = length_minus;
 648     }
 649   }
 650   *ctrl = slow_control;
 651   if (!(*ctrl)->is_top()) {
 652     Node* local_ctrl = *ctrl, *local_io = slow_i_o;
 653     MergeMemNode* local_mem = MergeMemNode::make(mem);
 654     transform_later(local_mem);
 655 
 656     // Generate the slow path, if needed.
 657     local_mem->set_memory_at(alias_idx, slow_mem);
 658 
 659     if (dest_uninitialized) {
 660       generate_clear_array(local_ctrl, local_mem,
 661                            adr_type, dest, basic_elem_type,
 662                            intcon(0), NULL,
 663                            alloc->in(AllocateNode::AllocSize));
 664     }
 665 
 666     local_mem = generate_slow_arraycopy(ac,
 667                                         &local_ctrl, local_mem, &local_io,
 668                                         adr_type,
 669                                         src, src_offset, dest, dest_offset,
 670                                         copy_length, /*dest_uninitialized*/false);
 671 
 672     result_region->init_req(slow_call_path, local_ctrl);
 673     result_i_o   ->init_req(slow_call_path, local_io);
 674     result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
 675   } else {
 676     ShouldNotReachHere(); // no call to generate_slow_arraycopy:
 677                           // projections were not extracted
 678   }
 679 
 680   // Remove unused edges.
 681   for (uint i = 1; i < result_region->req(); i++) {
 682     if (result_region->in(i) == NULL) {
 683       result_region->init_req(i, top());
 684     }
 685   }
 686 
 687   // Finished; return the combined state.
 688   *ctrl = result_region;
 689   *io = result_i_o;
 690   mem->set_memory_at(alias_idx, result_memory);
 691 
 692   // mem no longer guaranteed to stay a MergeMemNode
 693   Node* out_mem = mem;
 694   DEBUG_ONLY(mem = NULL);
 695 
 696   // The memory edges above are precise in order to model effects around
 697   // array copies accurately to allow value numbering of field loads around
 698   // arraycopy.  Such field loads, both before and after, are common in Java
 699   // collections and similar classes involving header/array data structures.
 700   //
 701   // But with low number of register or when some registers are used or killed
 702   // by arraycopy calls it causes registers spilling on stack. See 6544710.
 703   // The next memory barrier is added to avoid it. If the arraycopy can be
 704   // optimized away (which it can, sometimes) then we can manually remove
 705   // the membar also.
 706   //
 707   // Do not let reads from the cloned object float above the arraycopy.
 708   if (alloc != NULL && !alloc->initialization()->does_not_escape()) {
 709     // Do not let stores that initialize this object be reordered with
 710     // a subsequent store that would make this object accessible by
 711     // other threads.
 712     insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore);
 713   } else if (InsertMemBarAfterArraycopy) {
 714     insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
 715   }
 716 
 717   _igvn.replace_node(_memproj_fallthrough, out_mem);
 718   _igvn.replace_node(_ioproj_fallthrough, *io);
 719   _igvn.replace_node(_fallthroughcatchproj, *ctrl);
 720 
 721   return out_mem;
 722 }
 723 
 724 // Helper for initialization of arrays, creating a ClearArray.
 725 // It writes zero bits in [start..end), within the body of an array object.
 726 // The memory effects are all chained onto the 'adr_type' alias category.
 727 //
 728 // Since the object is otherwise uninitialized, we are free
 729 // to put a little "slop" around the edges of the cleared area,
 730 // as long as it does not go back into the array's header,
 731 // or beyond the array end within the heap.
 732 //
 733 // The lower edge can be rounded down to the nearest jint and the
 734 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
 735 //
 736 // Arguments:
 737 //   adr_type           memory slice where writes are generated
 738 //   dest               oop of the destination array
 739 //   basic_elem_type    element type of the destination
 740 //   slice_idx          array index of first element to store
 741 //   slice_len          number of elements to store (or NULL)
 742 //   dest_size          total size in bytes of the array object
 743 //
 744 // Exactly one of slice_len or dest_size must be non-NULL.
 745 // If dest_size is non-NULL, zeroing extends to the end of the object.
 746 // If slice_len is non-NULL, the slice_idx value must be a constant.
 747 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
 748                                             const TypePtr* adr_type,
 749                                             Node* dest,
 750                                             BasicType basic_elem_type,
 751                                             Node* slice_idx,
 752                                             Node* slice_len,
 753                                             Node* dest_size) {
 754   // one or the other but not both of slice_len and dest_size:
 755   assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
 756   if (slice_len == NULL)  slice_len = top();
 757   if (dest_size == NULL)  dest_size = top();
 758 
 759   uint alias_idx = C->get_alias_index(adr_type);
 760 
 761   // operate on this memory slice:
 762   Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
 763 
 764   // scaling and rounding of indexes:
 765   int scale = exact_log2(type2aelembytes(basic_elem_type));
 766   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 767   int clear_low = (-1 << scale) & (BytesPerInt  - 1);
 768   int bump_bit  = (-1 << scale) & BytesPerInt;
 769 
 770   // determine constant starts and ends
 771   const intptr_t BIG_NEG = -128;
 772   assert(BIG_NEG + 2*abase < 0, "neg enough");
 773   intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
 774   intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
 775   if (slice_len_con == 0) {
 776     return;                     // nothing to do here
 777   }
 778   intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
 779   intptr_t end_con   = _igvn.find_intptr_t_con(dest_size, -1);
 780   if (slice_idx_con >= 0 && slice_len_con >= 0) {
 781     assert(end_con < 0, "not two cons");
 782     end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
 783                        BytesPerLong);
 784   }
 785 
 786   if (start_con >= 0 && end_con >= 0) {
 787     // Constant start and end.  Simple.
 788     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 789                                        start_con, end_con, &_igvn);
 790   } else if (start_con >= 0 && dest_size != top()) {
 791     // Constant start, pre-rounded end after the tail of the array.
 792     Node* end = dest_size;
 793     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 794                                        start_con, end, &_igvn);
 795   } else if (start_con >= 0 && slice_len != top()) {
 796     // Constant start, non-constant end.  End needs rounding up.
 797     // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
 798     intptr_t end_base  = abase + (slice_idx_con << scale);
 799     int      end_round = (-1 << scale) & (BytesPerLong  - 1);
 800     Node*    end       = ConvI2X(slice_len);
 801     if (scale != 0)
 802       end = transform_later(new LShiftXNode(end, intcon(scale) ));
 803     end_base += end_round;
 804     end = transform_later(new AddXNode(end, MakeConX(end_base)) );
 805     end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
 806     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 807                                        start_con, end, &_igvn);
 808   } else if (start_con < 0 && dest_size != top()) {
 809     // Non-constant start, pre-rounded end after the tail of the array.
 810     // This is almost certainly a "round-to-end" operation.
 811     Node* start = slice_idx;
 812     start = ConvI2X(start);
 813     if (scale != 0)
 814       start = transform_later(new LShiftXNode( start, intcon(scale) ));
 815     start = transform_later(new AddXNode(start, MakeConX(abase)) );
 816     if ((bump_bit | clear_low) != 0) {
 817       int to_clear = (bump_bit | clear_low);
 818       // Align up mod 8, then store a jint zero unconditionally
 819       // just before the mod-8 boundary.
 820       if (((abase + bump_bit) & ~to_clear) - bump_bit
 821           < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
 822         bump_bit = 0;
 823         assert((abase & to_clear) == 0, "array base must be long-aligned");
 824       } else {
 825         // Bump 'start' up to (or past) the next jint boundary:
 826         start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
 827         assert((abase & clear_low) == 0, "array base must be int-aligned");
 828       }
 829       // Round bumped 'start' down to jlong boundary in body of array.
 830       start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
 831       if (bump_bit != 0) {
 832         // Store a zero to the immediately preceding jint:
 833         Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
 834         Node* p1 = basic_plus_adr(dest, x1);
 835         mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
 836         mem = transform_later(mem);
 837       }
 838     }
 839     Node* end = dest_size; // pre-rounded
 840     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 841                                        start, end, &_igvn);
 842   } else {
 843     // Non-constant start, unrounded non-constant end.
 844     // (Nobody zeroes a random midsection of an array using this routine.)
 845     ShouldNotReachHere();       // fix caller
 846   }
 847 
 848   // Done.
 849   merge_mem->set_memory_at(alias_idx, mem);
 850 }
 851 
 852 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
 853                                                 const TypePtr* adr_type,
 854                                                 BasicType basic_elem_type,
 855                                                 AllocateNode* alloc,
 856                                                 Node* src,  Node* src_offset,
 857                                                 Node* dest, Node* dest_offset,
 858                                                 Node* dest_size, bool dest_uninitialized) {
 859   // See if there is an advantage from block transfer.
 860   int scale = exact_log2(type2aelembytes(basic_elem_type));
 861   if (scale >= LogBytesPerLong)
 862     return false;               // it is already a block transfer
 863 
 864   // Look at the alignment of the starting offsets.
 865   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 866 
 867   intptr_t src_off_con  = (intptr_t) _igvn.find_int_con(src_offset, -1);
 868   intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
 869   if (src_off_con < 0 || dest_off_con < 0) {
 870     // At present, we can only understand constants.
 871     return false;
 872   }
 873 
 874   intptr_t src_off  = abase + (src_off_con  << scale);
 875   intptr_t dest_off = abase + (dest_off_con << scale);
 876 
 877   if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
 878     // Non-aligned; too bad.
 879     // One more chance:  Pick off an initial 32-bit word.
 880     // This is a common case, since abase can be odd mod 8.
 881     if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
 882         ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
 883       Node* sptr = basic_plus_adr(src,  src_off);
 884       Node* dptr = basic_plus_adr(dest, dest_off);
 885       uint alias_idx = C->get_alias_index(adr_type);
 886       bool is_mismatched = (basic_elem_type != T_INT);
 887       Node* sval = transform_later(
 888           LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(alias_idx), sptr, adr_type,
 889                          TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
 890                          false /*unaligned*/, is_mismatched));
 891       Node* st = transform_later(
 892           StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(alias_idx), dptr, adr_type,
 893                           sval, T_INT, MemNode::unordered));
 894       (*mem)->set_memory_at(alias_idx, st);
 895       src_off += BytesPerInt;
 896       dest_off += BytesPerInt;
 897     } else {
 898       return false;
 899     }
 900   }
 901   assert(src_off % BytesPerLong == 0, "");
 902   assert(dest_off % BytesPerLong == 0, "");
 903 
 904   // Do this copy by giant steps.
 905   Node* sptr  = basic_plus_adr(src,  src_off);
 906   Node* dptr  = basic_plus_adr(dest, dest_off);
 907   Node* countx = dest_size;
 908   countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
 909   countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
 910 
 911   bool disjoint_bases = true;   // since alloc != NULL
 912   generate_unchecked_arraycopy(ctrl, mem,
 913                                adr_type, T_LONG, disjoint_bases,
 914                                sptr, NULL, dptr, NULL, countx, dest_uninitialized);
 915 
 916   return true;
 917 }
 918 
 919 // Helper function; generates code for the slow case.
 920 // We make a call to a runtime method which emulates the native method,
 921 // but without the native wrapper overhead.
 922 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
 923                                                         Node** ctrl, Node* mem, Node** io,
 924                                                         const TypePtr* adr_type,
 925                                                         Node* src,  Node* src_offset,
 926                                                         Node* dest, Node* dest_offset,
 927                                                         Node* copy_length, bool dest_uninitialized) {
 928   assert(!dest_uninitialized, "Invariant");
 929 
 930   const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
 931   CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
 932                                           "slow_arraycopy",
 933                                           ac->jvms()->bci(), TypePtr::BOTTOM);
 934 
 935   call->init_req(TypeFunc::Control, *ctrl);
 936   call->init_req(TypeFunc::I_O    , *io);
 937   call->init_req(TypeFunc::Memory , mem);
 938   call->init_req(TypeFunc::ReturnAdr, top());
 939   call->init_req(TypeFunc::FramePtr, top());
 940   call->init_req(TypeFunc::Parms+0, src);
 941   call->init_req(TypeFunc::Parms+1, src_offset);
 942   call->init_req(TypeFunc::Parms+2, dest);
 943   call->init_req(TypeFunc::Parms+3, dest_offset);
 944   call->init_req(TypeFunc::Parms+4, copy_length);
 945   copy_call_debug_info(ac, call);
 946 
 947   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
 948   _igvn.replace_node(ac, call);
 949   transform_later(call);
 950 
 951   extract_call_projections(call);
 952   *ctrl = _fallthroughcatchproj->clone();
 953   transform_later(*ctrl);
 954 
 955   Node* m = _memproj_fallthrough->clone();
 956   transform_later(m);
 957 
 958   uint alias_idx = C->get_alias_index(adr_type);
 959   MergeMemNode* out_mem;
 960   if (alias_idx != Compile::AliasIdxBot) {
 961     out_mem = MergeMemNode::make(mem);
 962     out_mem->set_memory_at(alias_idx, m);
 963   } else {
 964     out_mem = MergeMemNode::make(m);
 965   }
 966   transform_later(out_mem);
 967 
 968   *io = _ioproj_fallthrough->clone();
 969   transform_later(*io);
 970 
 971   return out_mem;
 972 }
 973 
 974 // Helper function; generates code for cases requiring runtime checks.
 975 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
 976                                                      const TypePtr* adr_type,
 977                                                      Node* dest_elem_klass,
 978                                                      Node* src,  Node* src_offset,
 979                                                      Node* dest, Node* dest_offset,
 980                                                      Node* copy_length, bool dest_uninitialized) {
 981   if ((*ctrl)->is_top())  return NULL;
 982 
 983   address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
 984   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
 985     return NULL;
 986   }
 987 
 988   // Pick out the parameters required to perform a store-check
 989   // for the target array.  This is an optimistic check.  It will
 990   // look in each non-null element's class, at the desired klass's
 991   // super_check_offset, for the desired klass.
 992   int sco_offset = in_bytes(Klass::super_check_offset_offset());
 993   Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
 994   Node* n3 = new LoadINode(NULL, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
 995   Node* check_offset = ConvI2X(transform_later(n3));
 996   Node* check_value  = dest_elem_klass;
 997 
 998   Node* src_start  = array_element_address(src,  src_offset,  T_OBJECT);
 999   Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1000 
1001   const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1002   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1003                               src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1004 
1005   finish_arraycopy_call(call, ctrl, mem, adr_type);
1006 
1007   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1008   transform_later(proj);
1009 
1010   return proj;
1011 }
1012 
1013 // Helper function; generates code for cases requiring runtime checks.
1014 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1015                                                    const TypePtr* adr_type,
1016                                                    Node* src,  Node* src_offset,
1017                                                    Node* dest, Node* dest_offset,
1018                                                    Node* copy_length, bool dest_uninitialized) {
1019   if ((*ctrl)->is_top()) return NULL;
1020   assert(!dest_uninitialized, "Invariant");
1021 
1022   address copyfunc_addr = StubRoutines::generic_arraycopy();
1023   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1024     return NULL;
1025   }
1026 
1027   const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1028   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1029                               src, src_offset, dest, dest_offset, copy_length);
1030 
1031   finish_arraycopy_call(call, ctrl, mem, adr_type);
1032 
1033   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1034   transform_later(proj);
1035 
1036   return proj;
1037 }
1038 
1039 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1040 void PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1041                                                     const TypePtr* adr_type,
1042                                                     BasicType basic_elem_type,
1043                                                     bool disjoint_bases,
1044                                                     Node* src,  Node* src_offset,
1045                                                     Node* dest, Node* dest_offset,
1046                                                     Node* copy_length, bool dest_uninitialized) {
1047   if ((*ctrl)->is_top()) return;
1048 
1049   Node* src_start  = src;
1050   Node* dest_start = dest;
1051   if (src_offset != NULL || dest_offset != NULL) {
1052     src_start =  array_element_address(src, src_offset, basic_elem_type);
1053     dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1054   }
1055 
1056   // Figure out which arraycopy runtime method to call.
1057   const char* copyfunc_name = "arraycopy";
1058   address     copyfunc_addr =
1059       basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1060                           disjoint_bases, copyfunc_name, dest_uninitialized);
1061 
1062   const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1063   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1064                               src_start, dest_start, copy_length XTOP);
1065 
1066   finish_arraycopy_call(call, ctrl, mem, adr_type);
1067 }
1068 
1069 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1070   Node* ctrl = ac->in(TypeFunc::Control);
1071   Node* io = ac->in(TypeFunc::I_O);
1072   Node* src = ac->in(ArrayCopyNode::Src);
1073   Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1074   Node* dest = ac->in(ArrayCopyNode::Dest);
1075   Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1076   Node* length = ac->in(ArrayCopyNode::Length);
1077   MergeMemNode* merge_mem = NULL;
1078 
1079   if (ac->is_clonebasic()) {
1080     assert (src_offset == NULL && dest_offset == NULL, "for clone offsets should be null");
1081     Node* mem = ac->in(TypeFunc::Memory);
1082     const char* copyfunc_name = "arraycopy";
1083     address     copyfunc_addr =
1084       basictype2arraycopy(T_LONG, NULL, NULL,
1085                           true, copyfunc_name, true);
1086 
1087     const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
1088     const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1089 
1090     Node* call = make_leaf_call(ctrl, mem, call_type, copyfunc_addr, copyfunc_name, raw_adr_type, src, dest, length XTOP);
1091     transform_later(call);
1092 
1093     _igvn.replace_node(ac, call);
1094     return;
1095   } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_cloneoop()) {
1096     Node* mem = ac->in(TypeFunc::Memory);
1097     merge_mem = MergeMemNode::make(mem);
1098     transform_later(merge_mem);
1099 
1100     RegionNode* slow_region = new RegionNode(1);
1101     transform_later(slow_region);
1102 
1103     AllocateArrayNode* alloc = NULL;
1104     if (ac->is_alloc_tightly_coupled()) {
1105       alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1106       assert(alloc != NULL, "expect alloc");
1107     }
1108 
1109     const TypePtr* adr_type = _igvn.type(dest)->is_oopptr()->add_offset(Type::OffsetBot);
1110     if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1111       adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1112     }
1113     if (ac->_src_type != ac->_dest_type) {
1114       adr_type = TypeRawPtr::BOTTOM;
1115     }
1116     generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1117                        adr_type, T_OBJECT,
1118                        src, src_offset, dest, dest_offset, length,
1119                        true, !ac->is_copyofrange());
1120 
1121     return;
1122   }
1123 
1124   AllocateArrayNode* alloc = NULL;
1125   if (ac->is_alloc_tightly_coupled()) {
1126     alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1127     assert(alloc != NULL, "expect alloc");
1128   }
1129 
1130   assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1131 
1132   // Compile time checks.  If any of these checks cannot be verified at compile time,
1133   // we do not make a fast path for this call.  Instead, we let the call remain as it
1134   // is.  The checks we choose to mandate at compile time are:
1135   //
1136   // (1) src and dest are arrays.
1137   const Type* src_type = src->Value(&_igvn);
1138   const Type* dest_type = dest->Value(&_igvn);
1139   const TypeAryPtr* top_src = src_type->isa_aryptr();
1140   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1141 
1142   if (top_src  == NULL || top_src->klass()  == NULL ||
1143       top_dest == NULL || top_dest->klass() == NULL) {
1144     // Conservatively insert a memory barrier on all memory slices.
1145     // Do not let writes into the source float below the arraycopy.
1146     {
1147       Node* mem = ac->in(TypeFunc::Memory);
1148       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1149 
1150       merge_mem = MergeMemNode::make(mem);
1151       transform_later(merge_mem);
1152     }
1153 
1154     // Call StubRoutines::generic_arraycopy stub.
1155     Node* mem = generate_arraycopy(ac, NULL, &ctrl, merge_mem, &io,
1156                                    TypeRawPtr::BOTTOM, T_CONFLICT,
1157                                    src, src_offset, dest, dest_offset, length);
1158 
1159     // Do not let reads from the destination float above the arraycopy.
1160     // Since we cannot type the arrays, we don't know which slices
1161     // might be affected.  We could restrict this barrier only to those
1162     // memory slices which pertain to array elements--but don't bother.
1163     if (!InsertMemBarAfterArraycopy) {
1164       // (If InsertMemBarAfterArraycopy, there is already one in place.)
1165       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1166     }
1167     return;
1168   }
1169   // (2) src and dest arrays must have elements of the same BasicType
1170   // Figure out the size and type of the elements we will be copying.
1171   BasicType src_elem  = top_src->klass()->as_array_klass()->element_type()->basic_type();
1172   BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
1173   if (src_elem  == T_ARRAY)  src_elem  = T_OBJECT;
1174   if (dest_elem == T_ARRAY)  dest_elem = T_OBJECT;
1175 
1176   if (src_elem != dest_elem || dest_elem == T_VOID) {
1177     // The component types are not the same or are not recognized.  Punt.
1178     // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1179     {
1180       Node* mem = ac->in(TypeFunc::Memory);
1181       merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1182     }
1183 
1184     _igvn.replace_node(_memproj_fallthrough, merge_mem);
1185     _igvn.replace_node(_ioproj_fallthrough, io);
1186     _igvn.replace_node(_fallthroughcatchproj, ctrl);
1187     return;
1188   }
1189 
1190   //---------------------------------------------------------------------------
1191   // We will make a fast path for this call to arraycopy.
1192 
1193   // We have the following tests left to perform:
1194   //
1195   // (3) src and dest must not be null.
1196   // (4) src_offset must not be negative.
1197   // (5) dest_offset must not be negative.
1198   // (6) length must not be negative.
1199   // (7) src_offset + length must not exceed length of src.
1200   // (8) dest_offset + length must not exceed length of dest.
1201   // (9) each element of an oop array must be assignable
1202 
1203   {
1204     Node* mem = ac->in(TypeFunc::Memory);
1205     merge_mem = MergeMemNode::make(mem);
1206     transform_later(merge_mem);
1207   }
1208 
1209   RegionNode* slow_region = new RegionNode(1);
1210   transform_later(slow_region);
1211 
1212   if (!ac->is_arraycopy_validated()) {
1213     // (3) operands must not be null
1214     // We currently perform our null checks with the null_check routine.
1215     // This means that the null exceptions will be reported in the caller
1216     // rather than (correctly) reported inside of the native arraycopy call.
1217     // This should be corrected, given time.  We do our null check with the
1218     // stack pointer restored.
1219     // null checks done library_call.cpp
1220 
1221     // (4) src_offset must not be negative.
1222     generate_negative_guard(&ctrl, src_offset, slow_region);
1223 
1224     // (5) dest_offset must not be negative.
1225     generate_negative_guard(&ctrl, dest_offset, slow_region);
1226 
1227     // (6) length must not be negative (moved to generate_arraycopy()).
1228     // generate_negative_guard(length, slow_region);
1229 
1230     // (7) src_offset + length must not exceed length of src.
1231     Node* alen = ac->in(ArrayCopyNode::SrcLen);
1232     assert(alen != NULL, "need src len");
1233     generate_limit_guard(&ctrl,
1234                          src_offset, length,
1235                          alen,
1236                          slow_region);
1237 
1238     // (8) dest_offset + length must not exceed length of dest.
1239     alen = ac->in(ArrayCopyNode::DestLen);
1240     assert(alen != NULL, "need dest len");
1241     generate_limit_guard(&ctrl,
1242                          dest_offset, length,
1243                          alen,
1244                          slow_region);
1245 
1246     // (9) each element of an oop array must be assignable
1247     // The generate_arraycopy subroutine checks this.
1248   }
1249   // This is where the memory effects are placed:
1250   const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1251   if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1252     adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1253   }
1254   if (ac->_src_type != ac->_dest_type) {
1255     adr_type = TypeRawPtr::BOTTOM;
1256   }
1257 
1258   generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1259                      adr_type, dest_elem,
1260                      src, src_offset, dest, dest_offset, length,
1261                      false, false, slow_region);
1262 }