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 "gc/shared/barrierSet.hpp"
  27 #include "opto/arraycopynode.hpp"
  28 #include "oops/objArrayKlass.hpp"
  29 #include "opto/convertnode.hpp"
  30 #include "opto/graphKit.hpp"
  31 #include "opto/macro.hpp"
  32 #include "opto/runtime.hpp"
  33 #include "utilities/align.hpp"
  34 
  35 
  36 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, Node* precedent) {
  37   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
  38   mb->init_req(TypeFunc::Control, *ctrl);
  39   mb->init_req(TypeFunc::Memory, *mem);
  40   transform_later(mb);
  41   *ctrl = new ProjNode(mb,TypeFunc::Control);
  42   transform_later(*ctrl);
  43   Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
  44   transform_later(mem_proj);
  45   *mem = mem_proj;
  46 }
  47 
  48 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
  49   uint shift  = exact_log2(type2aelembytes(elembt));
  50   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
  51   Node* base =  basic_plus_adr(ary, header);
  52 #ifdef _LP64
  53   // see comment in GraphKit::array_element_address
  54   int index_max = max_jint - 1;  // array size is max_jint, index is one less
  55   const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
  56   idx = transform_later( new ConvI2LNode(idx, lidxtype) );
  57 #endif
  58   Node* scale = new LShiftXNode(idx, intcon(shift));
  59   transform_later(scale);
  60   return basic_plus_adr(ary, base, scale);
  61 }
  62 
  63 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
  64   return transform_later(new ConvI2LNode(offset));
  65 }
  66 
  67 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
  68                                        const TypeFunc* call_type, address call_addr,
  69                                        const char* call_name,
  70                                        const TypePtr* adr_type,
  71                                        Node* parm0, Node* parm1,
  72                                        Node* parm2, Node* parm3,
  73                                        Node* parm4, Node* parm5,
  74                                        Node* parm6, Node* parm7) {
  75   Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
  76   call->init_req(TypeFunc::Control, ctrl);
  77   call->init_req(TypeFunc::I_O    , top());
  78   call->init_req(TypeFunc::Memory , mem);
  79   call->init_req(TypeFunc::ReturnAdr, top());
  80   call->init_req(TypeFunc::FramePtr, top());
  81 
  82   // Hook each parm in order.  Stop looking at the first NULL.
  83   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
  84   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
  85   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
  86   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
  87   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
  88   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
  89   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
  90   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
  91     /* close each nested if ===> */  } } } } } } } }
  92   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
  93 
  94   return call;
  95 }
  96 
  97 
  98 //------------------------------generate_guard---------------------------
  99 // Helper function for generating guarded fast-slow graph structures.
 100 // The given 'test', if true, guards a slow path.  If the test fails
 101 // then a fast path can be taken.  (We generally hope it fails.)
 102 // In all cases, GraphKit::control() is updated to the fast path.
 103 // The returned value represents the control for the slow path.
 104 // The return value is never 'top'; it is either a valid control
 105 // or NULL if it is obvious that the slow path can never be taken.
 106 // Also, if region and the slow control are not NULL, the slow edge
 107 // is appended to the region.
 108 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
 109   if ((*ctrl)->is_top()) {
 110     // Already short circuited.
 111     return NULL;
 112   }
 113   // Build an if node and its projections.
 114   // If test is true we take the slow path, which we assume is uncommon.
 115   if (_igvn.type(test) == TypeInt::ZERO) {
 116     // The slow branch is never taken.  No need to build this guard.
 117     return NULL;
 118   }
 119 
 120   IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
 121   transform_later(iff);
 122 
 123   Node* if_slow = new IfTrueNode(iff);
 124   transform_later(if_slow);
 125 
 126   if (region != NULL) {
 127     region->add_req(if_slow);
 128   }
 129 
 130   Node* if_fast = new IfFalseNode(iff);
 131   transform_later(if_fast);
 132 
 133   *ctrl = if_fast;
 134 
 135   return if_slow;
 136 }
 137 
 138 inline Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
 139   return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
 140 }
 141 
 142 inline Node* PhaseMacroExpand::generate_fair_guard(Node** ctrl, Node* test, RegionNode* region) {
 143   return generate_guard(ctrl, test, region, PROB_FAIR);
 144 }
 145 
 146 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
 147   if ((*ctrl)->is_top())
 148     return;                // already stopped
 149   if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
 150     return;                // index is already adequately typed
 151   Node* cmp_lt = new CmpINode(index, intcon(0));
 152   transform_later(cmp_lt);
 153   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 154   transform_later(bol_lt);
 155   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 156 }
 157 
 158 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
 159   if ((*ctrl)->is_top())
 160     return;                // already stopped
 161   bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
 162   if (zero_offset && subseq_length->eqv_uncast(array_length))
 163     return;                // common case of whole-array copy
 164   Node* last = subseq_length;
 165   if (!zero_offset) {            // last += offset
 166     last = new AddINode(last, offset);
 167     transform_later(last);
 168   }
 169   Node* cmp_lt = new CmpUNode(array_length, last);
 170   transform_later(cmp_lt);
 171   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 172   transform_later(bol_lt);
 173   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 174 }
 175 
 176 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
 177   if ((*ctrl)->is_top())  return NULL;
 178 
 179   if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
 180     return NULL;                // index is already adequately typed
 181   Node* cmp_le = new CmpINode(index, intcon(0));
 182   transform_later(cmp_le);
 183   BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
 184   Node* bol_le = new BoolNode(cmp_le, le_or_eq);
 185   transform_later(bol_le);
 186   Node* is_notp = generate_guard(ctrl, bol_le, NULL, PROB_MIN);
 187 
 188   return is_notp;
 189 }
 190 
 191 Node* PhaseMacroExpand::generate_flattened_array_guard(Node** ctrl, Node* mem, Node* obj_or_klass, RegionNode* region) {
 192   return generate_array_guard(ctrl, mem, obj_or_klass, region, Klass::_lh_array_tag_vt_value);
 193 }
 194 
 195 Node* PhaseMacroExpand::generate_object_array_guard(Node** ctrl, Node* mem, Node* obj_or_klass, RegionNode* region) {
 196   return generate_array_guard(ctrl, mem, obj_or_klass, region, Klass::_lh_array_tag_obj_value);
 197 }
 198 
 199 Node* PhaseMacroExpand::generate_array_guard(Node** ctrl, Node* mem, Node* obj_or_klass, RegionNode* region, jint lh_con) {
 200   if ((*ctrl)->is_top())  return NULL;
 201 
 202   Node* kls = NULL;
 203   if (_igvn.type(obj_or_klass)->isa_oopptr()) {
 204     Node* k_adr = basic_plus_adr(obj_or_klass, oopDesc::klass_offset_in_bytes());
 205     kls = transform_later(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
 206   } else {
 207     assert(_igvn.type(obj_or_klass)->isa_klassptr(), "what else?");
 208     kls = obj_or_klass;
 209   }
 210   Node* layout_val = make_load(NULL, mem, kls, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
 211 
 212   layout_val = transform_later(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
 213   Node* cmp = transform_later(new CmpINode(layout_val, intcon(lh_con)));
 214   Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
 215 
 216   return generate_fair_guard(ctrl, bol, region);
 217 }
 218 
 219 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
 220   transform_later(call);
 221 
 222   *ctrl = new ProjNode(call,TypeFunc::Control);
 223   transform_later(*ctrl);
 224   Node* newmem = new ProjNode(call, TypeFunc::Memory);
 225   transform_later(newmem);
 226 
 227   uint alias_idx = C->get_alias_index(adr_type);
 228   if (alias_idx != Compile::AliasIdxBot) {
 229     *mem = MergeMemNode::make(*mem);
 230     (*mem)->set_memory_at(alias_idx, newmem);
 231   } else {
 232     *mem = MergeMemNode::make(newmem);
 233   }
 234   transform_later(*mem);
 235 }
 236 
 237 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
 238                                               Node* src_offset,
 239                                               Node* dest_offset,
 240                                               bool disjoint_bases,
 241                                               const char* &name,
 242                                               bool dest_uninitialized) {
 243   const TypeInt* src_offset_inttype  = _igvn.find_int_type(src_offset);;
 244   const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);;
 245 
 246   bool aligned = false;
 247   bool disjoint = disjoint_bases;
 248 
 249   // if the offsets are the same, we can treat the memory regions as
 250   // disjoint, because either the memory regions are in different arrays,
 251   // or they are identical (which we can treat as disjoint.)  We can also
 252   // treat a copy with a destination index  less that the source index
 253   // as disjoint since a low->high copy will work correctly in this case.
 254   if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
 255       dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
 256     // both indices are constants
 257     int s_offs = src_offset_inttype->get_con();
 258     int d_offs = dest_offset_inttype->get_con();
 259     int element_size = type2aelembytes(t);
 260     aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
 261               ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
 262     if (s_offs >= d_offs)  disjoint = true;
 263   } else if (src_offset == dest_offset && src_offset != NULL) {
 264     // This can occur if the offsets are identical non-constants.
 265     disjoint = true;
 266   }
 267 
 268   return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
 269 }
 270 
 271 bool PhaseMacroExpand::can_try_zeroing_elimination(AllocateArrayNode* alloc,
 272                                                    Node* src,
 273                                                    Node* dest) const {
 274   const TypeAryPtr* top_dest = _igvn.type(dest)->isa_aryptr();
 275 
 276   if (top_dest != NULL) {
 277     if (top_dest->klass() == NULL) {
 278       return false;
 279     }
 280   }
 281 
 282   return ReduceBulkZeroing
 283     && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
 284     && !src->eqv_uncast(dest)
 285     && alloc != NULL
 286     && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
 287     && alloc->maybe_set_complete(&_igvn);
 288 }
 289 
 290 #define XTOP LP64_ONLY(COMMA top())
 291 
 292 // Generate an optimized call to arraycopy.
 293 // Caller must guard against non-arrays.
 294 // Caller must determine a common array basic-type for both arrays.
 295 // Caller must validate offsets against array bounds.
 296 // The slow_region has already collected guard failure paths
 297 // (such as out of bounds length or non-conformable array types).
 298 // The generated code has this shape, in general:
 299 //
 300 //     if (length == 0)  return   // via zero_path
 301 //     slowval = -1
 302 //     if (types unknown) {
 303 //       slowval = call generic copy loop
 304 //       if (slowval == 0)  return  // via checked_path
 305 //     } else if (indexes in bounds) {
 306 //       if ((is object array) && !(array type check)) {
 307 //         slowval = call checked copy loop
 308 //         if (slowval == 0)  return  // via checked_path
 309 //       } else {
 310 //         call bulk copy loop
 311 //         return  // via fast_path
 312 //       }
 313 //     }
 314 //     // adjust params for remaining work:
 315 //     if (slowval != -1) {
 316 //       n = -1^slowval; src_offset += n; dest_offset += n; length -= n
 317 //     }
 318 //   slow_region:
 319 //     call slow arraycopy(src, src_offset, dest, dest_offset, length)
 320 //     return  // via slow_call_path
 321 //
 322 // This routine is used from several intrinsics:  System.arraycopy,
 323 // Object.clone (the array subcase), and Arrays.copyOf[Range].
 324 //
 325 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
 326                                            Node** ctrl, MergeMemNode* mem, Node** io,
 327                                            const TypePtr* adr_type,
 328                                            BasicType basic_elem_type,
 329                                            Node* src,  Node* src_offset,
 330                                            Node* dest, Node* dest_offset,
 331                                            Node* copy_length,
 332                                            Node* dest_length,
 333                                            bool disjoint_bases,
 334                                            bool length_never_negative,
 335                                            RegionNode* slow_region) {
 336   if (slow_region == NULL) {
 337     slow_region = new RegionNode(1);
 338     transform_later(slow_region);
 339   }
 340 
 341   bool  dest_uninitialized = false;
 342   Node* default_value = NULL;
 343   Node* raw_default_value = NULL;
 344 
 345   // See if this is the initialization of a newly-allocated array.
 346   // If so, we will take responsibility here for initializing it to zero.
 347   // (Note:  Because tightly_coupled_allocation performs checks on the
 348   // out-edges of the dest, we need to avoid making derived pointers
 349   // from it until we have checked its uses.)
 350   if (can_try_zeroing_elimination(alloc, src, dest) &&
 351       basic_elem_type != T_CONFLICT /* avoid corner case */) {
 352     // "You break it, you buy it."
 353     InitializeNode* init = alloc->initialization();
 354     assert(init->is_complete(), "we just did this");
 355     init->set_complete_with_arraycopy();
 356     assert(dest->is_CheckCastPP(), "sanity");
 357     assert(dest->in(0)->in(0) == init, "dest pinned");
 358     adr_type = TypeRawPtr::BOTTOM;  // all initializations are into raw memory
 359     // From this point on, every exit path is responsible for
 360     // initializing any non-copied parts of the object to zero.
 361     // Also, if this flag is set we make sure that arraycopy interacts properly
 362     // with G1, eliding pre-barriers. See CR 6627983.
 363     dest_uninitialized = true;
 364     default_value = alloc->in(AllocateNode::DefaultValue);
 365     raw_default_value = alloc->in(AllocateNode::RawDefaultValue);
 366   } else {
 367     // No zeroing elimination here.
 368     alloc             = NULL;
 369   }
 370 
 371   uint alias_idx = C->get_alias_index(adr_type);
 372 
 373   // Results are placed here:
 374   enum { fast_path        = 1,  // normal void-returning assembly stub
 375          checked_path     = 2,  // special assembly stub with cleanup
 376          slow_call_path   = 3,  // something went wrong; call the VM
 377          zero_path        = 4,  // bypass when length of copy is zero
 378          bcopy_path       = 5,  // copy primitive array by 64-bit blocks
 379          PATH_LIMIT       = 6
 380   };
 381   RegionNode* result_region = new RegionNode(PATH_LIMIT);
 382   PhiNode*    result_i_o    = new PhiNode(result_region, Type::ABIO);
 383   PhiNode*    result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
 384   assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
 385   transform_later(result_region);
 386   transform_later(result_i_o);
 387   transform_later(result_memory);
 388 
 389   // The slow_control path:
 390   Node* slow_control;
 391   Node* slow_i_o = *io;
 392   Node* slow_mem = mem->memory_at(alias_idx);
 393   DEBUG_ONLY(slow_control = (Node*) badAddress);
 394 
 395   // Checked control path:
 396   Node* checked_control = top();
 397   Node* checked_mem     = NULL;
 398   Node* checked_i_o     = NULL;
 399   Node* checked_value   = NULL;
 400 
 401   if (basic_elem_type == T_CONFLICT) {
 402     assert(!dest_uninitialized, "");
 403     Node* cv = generate_generic_arraycopy(ctrl, &mem,
 404                                           adr_type,
 405                                           src, src_offset, dest, dest_offset,
 406                                           copy_length, dest_uninitialized);
 407     if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
 408     checked_control = *ctrl;
 409     checked_i_o     = *io;
 410     checked_mem     = mem->memory_at(alias_idx);
 411     checked_value   = cv;
 412     *ctrl = top();
 413   }
 414 
 415   Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
 416   if (not_pos != NULL) {
 417     Node* local_ctrl = not_pos, *local_io = *io;
 418     MergeMemNode* local_mem = MergeMemNode::make(mem);
 419     transform_later(local_mem);
 420 
 421     // (6) length must not be negative.
 422     if (!length_never_negative) {
 423       generate_negative_guard(&local_ctrl, copy_length, slow_region);
 424     }
 425 
 426     // copy_length is 0.
 427     if (dest_uninitialized) {
 428       assert(!local_ctrl->is_top(), "no ctrl?");
 429       if (copy_length->eqv_uncast(dest_length)
 430           || _igvn.find_int_con(dest_length, 1) <= 0) {
 431         // There is no zeroing to do. No need for a secondary raw memory barrier.
 432       } else {
 433         // Clear the whole thing since there are no source elements to copy.
 434         generate_clear_array(local_ctrl, local_mem,
 435                              adr_type, dest,
 436                              default_value, raw_default_value,
 437                              basic_elem_type,
 438                              intcon(0), NULL,
 439                              alloc->in(AllocateNode::AllocSize));
 440         // Use a secondary InitializeNode as raw memory barrier.
 441         // Currently it is needed only on this path since other
 442         // paths have stub or runtime calls as raw memory barriers.
 443         MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
 444                                           Compile::AliasIdxRaw,
 445                                           top());
 446         transform_later(mb);
 447         mb->set_req(TypeFunc::Control,local_ctrl);
 448         mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
 449         local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
 450         local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
 451 
 452         InitializeNode* init = mb->as_Initialize();
 453         init->set_complete(&_igvn);  // (there is no corresponding AllocateNode)
 454       }
 455     }
 456 
 457     // Present the results of the fast call.
 458     result_region->init_req(zero_path, local_ctrl);
 459     result_i_o   ->init_req(zero_path, local_io);
 460     result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
 461   }
 462 
 463   if (!(*ctrl)->is_top() && dest_uninitialized) {
 464     // We have to initialize the *uncopied* part of the array to zero.
 465     // The copy destination is the slice dest[off..off+len].  The other slices
 466     // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
 467     Node* dest_size   = alloc->in(AllocateNode::AllocSize);
 468     Node* dest_tail   = transform_later( new AddINode(dest_offset, copy_length));
 469 
 470     // If there is a head section that needs zeroing, do it now.
 471     if (_igvn.find_int_con(dest_offset, -1) != 0) {
 472       generate_clear_array(*ctrl, mem,
 473                            adr_type, dest,
 474                            default_value, raw_default_value,
 475                            basic_elem_type,
 476                            intcon(0), dest_offset,
 477                            NULL);
 478     }
 479 
 480     // Next, perform a dynamic check on the tail length.
 481     // It is often zero, and we can win big if we prove this.
 482     // There are two wins:  Avoid generating the ClearArray
 483     // with its attendant messy index arithmetic, and upgrade
 484     // the copy to a more hardware-friendly word size of 64 bits.
 485     Node* tail_ctl = NULL;
 486     if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
 487       Node* cmp_lt   = transform_later( new CmpINode(dest_tail, dest_length) );
 488       Node* bol_lt   = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
 489       tail_ctl = generate_slow_guard(ctrl, bol_lt, NULL);
 490       assert(tail_ctl != NULL || !(*ctrl)->is_top(), "must be an outcome");
 491     }
 492 
 493     // At this point, let's assume there is no tail.
 494     if (!(*ctrl)->is_top() && alloc != NULL && basic_elem_type != T_OBJECT) {
 495       // There is no tail.  Try an upgrade to a 64-bit copy.
 496       bool didit = false;
 497       {
 498         Node* local_ctrl = *ctrl, *local_io = *io;
 499         MergeMemNode* local_mem = MergeMemNode::make(mem);
 500         transform_later(local_mem);
 501 
 502         didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
 503                                          adr_type, basic_elem_type, alloc,
 504                                          src, src_offset, dest, dest_offset,
 505                                          dest_size, dest_uninitialized);
 506         if (didit) {
 507           // Present the results of the block-copying fast call.
 508           result_region->init_req(bcopy_path, local_ctrl);
 509           result_i_o   ->init_req(bcopy_path, local_io);
 510           result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
 511         }
 512       }
 513       if (didit) {
 514         *ctrl = top();     // no regular fast path
 515       }
 516     }
 517 
 518     // Clear the tail, if any.
 519     if (tail_ctl != NULL) {
 520       Node* notail_ctl = (*ctrl)->is_top() ? NULL : *ctrl;
 521       *ctrl = tail_ctl;
 522       if (notail_ctl == NULL) {
 523         generate_clear_array(*ctrl, mem,
 524                              adr_type, dest,
 525                              default_value, raw_default_value,
 526                              basic_elem_type,
 527                              dest_tail, NULL,
 528                              dest_size);
 529       } else {
 530         // Make a local merge.
 531         Node* done_ctl = transform_later(new RegionNode(3));
 532         Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
 533         done_ctl->init_req(1, notail_ctl);
 534         done_mem->init_req(1, mem->memory_at(alias_idx));
 535         generate_clear_array(*ctrl, mem,
 536                              adr_type, dest,
 537                              default_value, raw_default_value,
 538                              basic_elem_type,
 539                              dest_tail, NULL,
 540                              dest_size);
 541         done_ctl->init_req(2, *ctrl);
 542         done_mem->init_req(2, mem->memory_at(alias_idx));
 543         *ctrl = done_ctl;
 544         mem->set_memory_at(alias_idx, done_mem);
 545       }
 546     }
 547   }
 548 
 549   BasicType copy_type = basic_elem_type;
 550   assert(basic_elem_type != T_ARRAY, "caller must fix this");
 551   if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
 552     // If src and dest have compatible element types, we can copy bits.
 553     // Types S[] and D[] are compatible if D is a supertype of S.
 554     //
 555     // If they are not, we will use checked_oop_disjoint_arraycopy,
 556     // which performs a fast optimistic per-oop check, and backs off
 557     // further to JVM_ArrayCopy on the first per-oop check that fails.
 558     // (Actually, we don't move raw bits only; the GC requires card marks.)
 559 
 560     // We don't need a subtype check for validated copies and Object[].clone()
 561     bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
 562                               ac->is_copyofrange_validated() || ac->is_cloneoop();
 563     if (!skip_subtype_check) {
 564       // Get the klass* for both src and dest
 565       Node* src_klass  = ac->in(ArrayCopyNode::SrcKlass);
 566       Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
 567 
 568       assert(src_klass != NULL && dest_klass != NULL, "should have klasses");
 569 
 570       // Generate the subtype check.
 571       // This might fold up statically, or then again it might not.
 572       //
 573       // Non-static example:  Copying List<String>.elements to a new String[].
 574       // The backing store for a List<String> is always an Object[],
 575       // but its elements are always type String, if the generic types
 576       // are correct at the source level.
 577       //
 578       // Test S[] against D[], not S against D, because (probably)
 579       // the secondary supertype cache is less busy for S[] than S.
 580       // This usually only matters when D is an interface.
 581       Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, &_igvn);
 582       // Plug failing path into checked_oop_disjoint_arraycopy
 583       if (not_subtype_ctrl != top()) {
 584         Node* local_ctrl = not_subtype_ctrl;
 585         MergeMemNode* local_mem = MergeMemNode::make(mem);
 586         transform_later(local_mem);
 587 
 588         // (At this point we can assume disjoint_bases, since types differ.)
 589         int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
 590         Node* p1 = basic_plus_adr(dest_klass, ek_offset);
 591         Node* n1 = LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
 592         Node* dest_elem_klass = transform_later(n1);
 593         Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
 594                                                 adr_type,
 595                                                 dest_elem_klass,
 596                                                 src, src_offset, dest, dest_offset,
 597                                                 ConvI2X(copy_length), dest_uninitialized);
 598         if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
 599         checked_control = local_ctrl;
 600         checked_i_o     = *io;
 601         checked_mem     = local_mem->memory_at(alias_idx);
 602         checked_value   = cv;
 603       }
 604     }
 605     // At this point we know we do not need type checks on oop stores.
 606 
 607     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 608     if (!bs->array_copy_requires_gc_barriers(alloc != NULL, copy_type, false, BarrierSetC2::Expansion)) {
 609       // If we do not need gc barriers, copy using the jint or jlong stub.
 610       copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
 611       assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
 612              "sizes agree");
 613     }
 614   }
 615 
 616   if (!(*ctrl)->is_top()) {
 617     // Generate the fast path, if possible.
 618     Node* local_ctrl = *ctrl;
 619     MergeMemNode* local_mem = MergeMemNode::make(mem);
 620     transform_later(local_mem);
 621 
 622     generate_unchecked_arraycopy(&local_ctrl, &local_mem,
 623                                  adr_type, copy_type, disjoint_bases,
 624                                  src, src_offset, dest, dest_offset,
 625                                  ConvI2X(copy_length), dest_uninitialized);
 626 
 627     // Present the results of the fast call.
 628     result_region->init_req(fast_path, local_ctrl);
 629     result_i_o   ->init_req(fast_path, *io);
 630     result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
 631   }
 632 
 633   // Here are all the slow paths up to this point, in one bundle:
 634   assert(slow_region != NULL, "allocated on entry");
 635   slow_control = slow_region;
 636   DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
 637 
 638   *ctrl = checked_control;
 639   if (!(*ctrl)->is_top()) {
 640     // Clean up after the checked call.
 641     // The returned value is either 0 or -1^K,
 642     // where K = number of partially transferred array elements.
 643     Node* cmp = new CmpINode(checked_value, intcon(0));
 644     transform_later(cmp);
 645     Node* bol = new BoolNode(cmp, BoolTest::eq);
 646     transform_later(bol);
 647     IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
 648     transform_later(iff);
 649 
 650     // If it is 0, we are done, so transfer to the end.
 651     Node* checks_done = new IfTrueNode(iff);
 652     transform_later(checks_done);
 653     result_region->init_req(checked_path, checks_done);
 654     result_i_o   ->init_req(checked_path, checked_i_o);
 655     result_memory->init_req(checked_path, checked_mem);
 656 
 657     // If it is not zero, merge into the slow call.
 658     *ctrl = new IfFalseNode(iff);
 659     transform_later(*ctrl);
 660     RegionNode* slow_reg2 = new RegionNode(3);
 661     PhiNode*    slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
 662     PhiNode*    slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
 663     transform_later(slow_reg2);
 664     transform_later(slow_i_o2);
 665     transform_later(slow_mem2);
 666     slow_reg2  ->init_req(1, slow_control);
 667     slow_i_o2  ->init_req(1, slow_i_o);
 668     slow_mem2  ->init_req(1, slow_mem);
 669     slow_reg2  ->init_req(2, *ctrl);
 670     slow_i_o2  ->init_req(2, checked_i_o);
 671     slow_mem2  ->init_req(2, checked_mem);
 672 
 673     slow_control = slow_reg2;
 674     slow_i_o     = slow_i_o2;
 675     slow_mem     = slow_mem2;
 676 
 677     if (alloc != NULL) {
 678       // We'll restart from the very beginning, after zeroing the whole thing.
 679       // This can cause double writes, but that's OK since dest is brand new.
 680       // So we ignore the low 31 bits of the value returned from the stub.
 681     } else {
 682       // We must continue the copy exactly where it failed, or else
 683       // another thread might see the wrong number of writes to dest.
 684       Node* checked_offset = new XorINode(checked_value, intcon(-1));
 685       Node* slow_offset    = new PhiNode(slow_reg2, TypeInt::INT);
 686       transform_later(checked_offset);
 687       transform_later(slow_offset);
 688       slow_offset->init_req(1, intcon(0));
 689       slow_offset->init_req(2, checked_offset);
 690 
 691       // Adjust the arguments by the conditionally incoming offset.
 692       Node* src_off_plus  = new AddINode(src_offset,  slow_offset);
 693       transform_later(src_off_plus);
 694       Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
 695       transform_later(dest_off_plus);
 696       Node* length_minus  = new SubINode(copy_length, slow_offset);
 697       transform_later(length_minus);
 698 
 699       // Tweak the node variables to adjust the code produced below:
 700       src_offset  = src_off_plus;
 701       dest_offset = dest_off_plus;
 702       copy_length = length_minus;
 703     }
 704   }
 705   *ctrl = slow_control;
 706   if (!(*ctrl)->is_top()) {
 707     Node* local_ctrl = *ctrl, *local_io = slow_i_o;
 708     MergeMemNode* local_mem = MergeMemNode::make(mem);
 709     transform_later(local_mem);
 710 
 711     // Generate the slow path, if needed.
 712     local_mem->set_memory_at(alias_idx, slow_mem);
 713 
 714     if (dest_uninitialized) {
 715       generate_clear_array(local_ctrl, local_mem,
 716                            adr_type, dest,
 717                            default_value, raw_default_value,
 718                            basic_elem_type,
 719                            intcon(0), NULL,
 720                            alloc->in(AllocateNode::AllocSize));
 721     }
 722 
 723     local_mem = generate_slow_arraycopy(ac,
 724                                         &local_ctrl, local_mem, &local_io,
 725                                         adr_type,
 726                                         src, src_offset, dest, dest_offset,
 727                                         copy_length, /*dest_uninitialized*/false);
 728 
 729     result_region->init_req(slow_call_path, local_ctrl);
 730     result_i_o   ->init_req(slow_call_path, local_io);
 731     result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
 732   } else {
 733     ShouldNotReachHere(); // no call to generate_slow_arraycopy:
 734                           // projections were not extracted
 735   }
 736 
 737   // Remove unused edges.
 738   for (uint i = 1; i < result_region->req(); i++) {
 739     if (result_region->in(i) == NULL) {
 740       result_region->init_req(i, top());
 741     }
 742   }
 743 
 744   // Finished; return the combined state.
 745   *ctrl = result_region;
 746   *io = result_i_o;
 747   mem->set_memory_at(alias_idx, result_memory);
 748 
 749   // mem no longer guaranteed to stay a MergeMemNode
 750   Node* out_mem = mem;
 751   DEBUG_ONLY(mem = NULL);
 752 
 753   // The memory edges above are precise in order to model effects around
 754   // array copies accurately to allow value numbering of field loads around
 755   // arraycopy.  Such field loads, both before and after, are common in Java
 756   // collections and similar classes involving header/array data structures.
 757   //
 758   // But with low number of register or when some registers are used or killed
 759   // by arraycopy calls it causes registers spilling on stack. See 6544710.
 760   // The next memory barrier is added to avoid it. If the arraycopy can be
 761   // optimized away (which it can, sometimes) then we can manually remove
 762   // the membar also.
 763   //
 764   // Do not let reads from the cloned object float above the arraycopy.
 765   if (alloc != NULL && !alloc->initialization()->does_not_escape()) {
 766     // Do not let stores that initialize this object be reordered with
 767     // a subsequent store that would make this object accessible by
 768     // other threads.
 769     insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore);
 770   } else if (InsertMemBarAfterArraycopy) {
 771     insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
 772   } else if (adr_type == TypeRawPtr::BOTTOM) {
 773     // Do not let reads from the destination float above the arraycopy.
 774     // Since we cannot type the arrays, we don't know which slices
 775     // might be affected.  We could restrict this barrier only to those
 776     // memory slices which pertain to array elements--but don't bother.
 777     insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
 778   }
 779 
 780   _igvn.replace_node(_memproj_fallthrough, out_mem);
 781   _igvn.replace_node(_ioproj_fallthrough, *io);
 782   _igvn.replace_node(_fallthroughcatchproj, *ctrl);
 783 
 784 #ifdef ASSERT
 785   const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
 786   if (dest_t->is_known_instance()) {
 787     ArrayCopyNode* ac = NULL;
 788     assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
 789     assert(ac == NULL, "no arraycopy anymore");
 790   }
 791 #endif
 792 
 793   return out_mem;
 794 }
 795 
 796 // Helper for initialization of arrays, creating a ClearArray.
 797 // It writes zero bits in [start..end), within the body of an array object.
 798 // The memory effects are all chained onto the 'adr_type' alias category.
 799 //
 800 // Since the object is otherwise uninitialized, we are free
 801 // to put a little "slop" around the edges of the cleared area,
 802 // as long as it does not go back into the array's header,
 803 // or beyond the array end within the heap.
 804 //
 805 // The lower edge can be rounded down to the nearest jint and the
 806 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
 807 //
 808 // Arguments:
 809 //   adr_type           memory slice where writes are generated
 810 //   dest               oop of the destination array
 811 //   basic_elem_type    element type of the destination
 812 //   slice_idx          array index of first element to store
 813 //   slice_len          number of elements to store (or NULL)
 814 //   dest_size          total size in bytes of the array object
 815 //
 816 // Exactly one of slice_len or dest_size must be non-NULL.
 817 // If dest_size is non-NULL, zeroing extends to the end of the object.
 818 // If slice_len is non-NULL, the slice_idx value must be a constant.
 819 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
 820                                             const TypePtr* adr_type,
 821                                             Node* dest,
 822                                             Node* val,
 823                                             Node* raw_val,
 824                                             BasicType basic_elem_type,
 825                                             Node* slice_idx,
 826                                             Node* slice_len,
 827                                             Node* dest_size) {
 828   // one or the other but not both of slice_len and dest_size:
 829   assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
 830   if (slice_len == NULL)  slice_len = top();
 831   if (dest_size == NULL)  dest_size = top();
 832 
 833   uint alias_idx = C->get_alias_index(adr_type);
 834 
 835   // operate on this memory slice:
 836   Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
 837 
 838   // scaling and rounding of indexes:
 839   assert(basic_elem_type != T_VALUETYPE, "should have been converted to a basic type copy");
 840   int scale = exact_log2(type2aelembytes(basic_elem_type));
 841   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 842   int clear_low = (-1 << scale) & (BytesPerInt  - 1);
 843   int bump_bit  = (-1 << scale) & BytesPerInt;
 844 
 845   // determine constant starts and ends
 846   const intptr_t BIG_NEG = -128;
 847   assert(BIG_NEG + 2*abase < 0, "neg enough");
 848   intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
 849   intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
 850   if (slice_len_con == 0) {
 851     return;                     // nothing to do here
 852   }
 853   intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
 854   intptr_t end_con   = _igvn.find_intptr_t_con(dest_size, -1);
 855   if (slice_idx_con >= 0 && slice_len_con >= 0) {
 856     assert(end_con < 0, "not two cons");
 857     end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
 858                        BytesPerLong);
 859   }
 860 
 861   if (start_con >= 0 && end_con >= 0) {
 862     // Constant start and end.  Simple.
 863     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
 864                                        start_con, end_con, &_igvn);
 865   } else if (start_con >= 0 && dest_size != top()) {
 866     // Constant start, pre-rounded end after the tail of the array.
 867     Node* end = dest_size;
 868     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
 869                                        start_con, end, &_igvn);
 870   } else if (start_con >= 0 && slice_len != top()) {
 871     // Constant start, non-constant end.  End needs rounding up.
 872     // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
 873     intptr_t end_base  = abase + (slice_idx_con << scale);
 874     int      end_round = (-1 << scale) & (BytesPerLong  - 1);
 875     Node*    end       = ConvI2X(slice_len);
 876     if (scale != 0)
 877       end = transform_later(new LShiftXNode(end, intcon(scale) ));
 878     end_base += end_round;
 879     end = transform_later(new AddXNode(end, MakeConX(end_base)) );
 880     end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
 881     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
 882                                        start_con, end, &_igvn);
 883   } else if (start_con < 0 && dest_size != top()) {
 884     // Non-constant start, pre-rounded end after the tail of the array.
 885     // This is almost certainly a "round-to-end" operation.
 886     Node* start = slice_idx;
 887     start = ConvI2X(start);
 888     if (scale != 0)
 889       start = transform_later(new LShiftXNode( start, intcon(scale) ));
 890     start = transform_later(new AddXNode(start, MakeConX(abase)) );
 891     if ((bump_bit | clear_low) != 0) {
 892       int to_clear = (bump_bit | clear_low);
 893       // Align up mod 8, then store a jint zero unconditionally
 894       // just before the mod-8 boundary.
 895       if (((abase + bump_bit) & ~to_clear) - bump_bit
 896           < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
 897         bump_bit = 0;
 898         assert((abase & to_clear) == 0, "array base must be long-aligned");
 899       } else {
 900         // Bump 'start' up to (or past) the next jint boundary:
 901         start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
 902         assert((abase & clear_low) == 0, "array base must be int-aligned");
 903       }
 904       // Round bumped 'start' down to jlong boundary in body of array.
 905       start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
 906       if (bump_bit != 0) {
 907         // Store a zero to the immediately preceding jint:
 908         Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
 909         Node* p1 = basic_plus_adr(dest, x1);
 910         if (val == NULL) {
 911           assert(raw_val == NULL, "val may not be null");
 912           mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
 913         } else {
 914           assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
 915           mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
 916         }
 917         mem = transform_later(mem);
 918       }
 919     }
 920     Node* end = dest_size; // pre-rounded
 921     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
 922                                        start, end, &_igvn);
 923   } else {
 924     // Non-constant start, unrounded non-constant end.
 925     // (Nobody zeroes a random midsection of an array using this routine.)
 926     ShouldNotReachHere();       // fix caller
 927   }
 928 
 929   // Done.
 930   merge_mem->set_memory_at(alias_idx, mem);
 931 }
 932 
 933 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
 934                                                 const TypePtr* adr_type,
 935                                                 BasicType basic_elem_type,
 936                                                 AllocateNode* alloc,
 937                                                 Node* src,  Node* src_offset,
 938                                                 Node* dest, Node* dest_offset,
 939                                                 Node* dest_size, bool dest_uninitialized) {
 940   // See if there is an advantage from block transfer.
 941   int scale = exact_log2(type2aelembytes(basic_elem_type));
 942   if (scale >= LogBytesPerLong)
 943     return false;               // it is already a block transfer
 944 
 945   // Look at the alignment of the starting offsets.
 946   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 947 
 948   intptr_t src_off_con  = (intptr_t) _igvn.find_int_con(src_offset, -1);
 949   intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
 950   if (src_off_con < 0 || dest_off_con < 0) {
 951     // At present, we can only understand constants.
 952     return false;
 953   }
 954 
 955   intptr_t src_off  = abase + (src_off_con  << scale);
 956   intptr_t dest_off = abase + (dest_off_con << scale);
 957 
 958   if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
 959     // Non-aligned; too bad.
 960     // One more chance:  Pick off an initial 32-bit word.
 961     // This is a common case, since abase can be odd mod 8.
 962     if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
 963         ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
 964       Node* sptr = basic_plus_adr(src,  src_off);
 965       Node* dptr = basic_plus_adr(dest, dest_off);
 966       const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
 967       assert(s_adr_type->isa_aryptr(), "impossible slice");
 968       uint s_alias_idx = C->get_alias_index(s_adr_type);
 969       uint d_alias_idx = C->get_alias_index(adr_type);
 970       bool is_mismatched = (basic_elem_type != T_INT);
 971       Node* sval = transform_later(
 972           LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
 973                          TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
 974                          false /*unaligned*/, is_mismatched));
 975       Node* st = transform_later(
 976           StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
 977                           sval, T_INT, MemNode::unordered));
 978       if (is_mismatched) {
 979         st->as_Store()->set_mismatched_access();
 980       }
 981       (*mem)->set_memory_at(d_alias_idx, st);
 982       src_off += BytesPerInt;
 983       dest_off += BytesPerInt;
 984     } else {
 985       return false;
 986     }
 987   }
 988   assert(src_off % BytesPerLong == 0, "");
 989   assert(dest_off % BytesPerLong == 0, "");
 990 
 991   // Do this copy by giant steps.
 992   Node* sptr  = basic_plus_adr(src,  src_off);
 993   Node* dptr  = basic_plus_adr(dest, dest_off);
 994   Node* countx = dest_size;
 995   countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
 996   countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
 997 
 998   bool disjoint_bases = true;   // since alloc != NULL
 999   generate_unchecked_arraycopy(ctrl, mem,
1000                                adr_type, T_LONG, disjoint_bases,
1001                                sptr, NULL, dptr, NULL, countx, dest_uninitialized);
1002 
1003   return true;
1004 }
1005 
1006 // Helper function; generates code for the slow case.
1007 // We make a call to a runtime method which emulates the native method,
1008 // but without the native wrapper overhead.
1009 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
1010                                                         Node** ctrl, Node* mem, Node** io,
1011                                                         const TypePtr* adr_type,
1012                                                         Node* src,  Node* src_offset,
1013                                                         Node* dest, Node* dest_offset,
1014                                                         Node* copy_length, bool dest_uninitialized) {
1015   assert(!dest_uninitialized, "Invariant");
1016 
1017   const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1018   CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1019                                           "slow_arraycopy",
1020                                           ac->jvms()->bci(), TypePtr::BOTTOM);
1021 
1022   call->init_req(TypeFunc::Control, *ctrl);
1023   call->init_req(TypeFunc::I_O    , *io);
1024   call->init_req(TypeFunc::Memory , mem);
1025   call->init_req(TypeFunc::ReturnAdr, top());
1026   call->init_req(TypeFunc::FramePtr, top());
1027   call->init_req(TypeFunc::Parms+0, src);
1028   call->init_req(TypeFunc::Parms+1, src_offset);
1029   call->init_req(TypeFunc::Parms+2, dest);
1030   call->init_req(TypeFunc::Parms+3, dest_offset);
1031   call->init_req(TypeFunc::Parms+4, copy_length);
1032   copy_call_debug_info(ac, call);
1033 
1034   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1035   _igvn.replace_node(ac, call);
1036   transform_later(call);
1037 
1038   extract_call_projections(call);
1039   *ctrl = _fallthroughcatchproj->clone();
1040   transform_later(*ctrl);
1041 
1042   Node* m = _memproj_fallthrough->clone();
1043   transform_later(m);
1044 
1045   uint alias_idx = C->get_alias_index(adr_type);
1046   MergeMemNode* out_mem;
1047   if (alias_idx != Compile::AliasIdxBot) {
1048     out_mem = MergeMemNode::make(mem);
1049     out_mem->set_memory_at(alias_idx, m);
1050   } else {
1051     out_mem = MergeMemNode::make(m);
1052   }
1053   transform_later(out_mem);
1054 
1055   *io = _ioproj_fallthrough->clone();
1056   transform_later(*io);
1057 
1058   return out_mem;
1059 }
1060 
1061 // Helper function; generates code for cases requiring runtime checks.
1062 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1063                                                      const TypePtr* adr_type,
1064                                                      Node* dest_elem_klass,
1065                                                      Node* src,  Node* src_offset,
1066                                                      Node* dest, Node* dest_offset,
1067                                                      Node* copy_length, bool dest_uninitialized) {
1068   if ((*ctrl)->is_top())  return NULL;
1069 
1070   address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1071   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1072     return NULL;
1073   }
1074 
1075   // Pick out the parameters required to perform a store-check
1076   // for the target array.  This is an optimistic check.  It will
1077   // look in each non-null element's class, at the desired klass's
1078   // super_check_offset, for the desired klass.
1079   int sco_offset = in_bytes(Klass::super_check_offset_offset());
1080   Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1081   Node* n3 = new LoadINode(NULL, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1082   Node* check_offset = ConvI2X(transform_later(n3));
1083   Node* check_value  = dest_elem_klass;
1084 
1085   Node* src_start  = array_element_address(src,  src_offset,  T_OBJECT);
1086   Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1087 
1088   const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1089   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1090                               src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1091 
1092   finish_arraycopy_call(call, ctrl, mem, adr_type);
1093 
1094   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1095   transform_later(proj);
1096 
1097   return proj;
1098 }
1099 
1100 // Helper function; generates code for cases requiring runtime checks.
1101 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1102                                                    const TypePtr* adr_type,
1103                                                    Node* src,  Node* src_offset,
1104                                                    Node* dest, Node* dest_offset,
1105                                                    Node* copy_length, bool dest_uninitialized) {
1106   if ((*ctrl)->is_top()) return NULL;
1107   assert(!dest_uninitialized, "Invariant");
1108 
1109   address copyfunc_addr = StubRoutines::generic_arraycopy();
1110   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1111     return NULL;
1112   }
1113 
1114   const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1115   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1116                               src, src_offset, dest, dest_offset, copy_length);
1117 
1118   finish_arraycopy_call(call, ctrl, mem, adr_type);
1119 
1120   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1121   transform_later(proj);
1122 
1123   return proj;
1124 }
1125 
1126 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1127 void PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1128                                                     const TypePtr* adr_type,
1129                                                     BasicType basic_elem_type,
1130                                                     bool disjoint_bases,
1131                                                     Node* src,  Node* src_offset,
1132                                                     Node* dest, Node* dest_offset,
1133                                                     Node* copy_length, bool dest_uninitialized) {
1134   if ((*ctrl)->is_top()) return;
1135 
1136   Node* src_start  = src;
1137   Node* dest_start = dest;
1138   if (src_offset != NULL || dest_offset != NULL) {
1139     src_start =  array_element_address(src, src_offset, basic_elem_type);
1140     dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1141   }
1142 
1143   // Figure out which arraycopy runtime method to call.
1144   const char* copyfunc_name = "arraycopy";
1145   address     copyfunc_addr =
1146       basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1147                           disjoint_bases, copyfunc_name, dest_uninitialized);
1148 
1149   const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1150   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1151                               src_start, dest_start, copy_length XTOP);
1152 
1153   finish_arraycopy_call(call, ctrl, mem, adr_type);
1154 }
1155 
1156 const TypePtr* PhaseMacroExpand::adjust_parameters_for_vt(const TypeAryPtr* top_dest, Node*& src_offset,
1157                                                           Node*& dest_offset, Node*& length, BasicType& dest_elem,
1158                                                           Node*& dest_length) {
1159   assert(top_dest->klass()->is_value_array_klass(), "inconsistent");
1160   int elem_size = ((ciValueArrayKlass*)top_dest->klass())->element_byte_size();
1161   if (elem_size >= 8) {
1162     if (elem_size > 8) {
1163       // treat as array of long but scale length, src offset and dest offset
1164       assert((elem_size % 8) == 0, "not a power of 2?");
1165       int factor = elem_size / 8;
1166       length = transform_later(new MulINode(length, intcon(factor)));
1167       src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1168       dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1169       if (dest_length != NULL) {
1170         dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1171       }
1172       elem_size = 8;
1173     }
1174     dest_elem = T_LONG;
1175   } else if (elem_size == 4) {
1176     dest_elem = T_INT;
1177   } else if (elem_size == 2) {
1178     dest_elem = T_CHAR;
1179   } else if (elem_size == 1) {
1180     dest_elem = T_BYTE;
1181   } else {
1182     ShouldNotReachHere();
1183   }
1184   return TypeRawPtr::BOTTOM;
1185 }
1186 
1187 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1188   Node* ctrl = ac->in(TypeFunc::Control);
1189   Node* io = ac->in(TypeFunc::I_O);
1190   Node* src = ac->in(ArrayCopyNode::Src);
1191   Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1192   Node* dest = ac->in(ArrayCopyNode::Dest);
1193   Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1194   Node* length = ac->in(ArrayCopyNode::Length);
1195   MergeMemNode* merge_mem = NULL;
1196 
1197   if (ac->is_clonebasic()) {
1198     assert (src_offset == NULL && dest_offset == NULL, "for clone offsets should be null");
1199     Node* mem = ac->in(TypeFunc::Memory);
1200     const char* copyfunc_name = "arraycopy";
1201     address     copyfunc_addr =
1202       basictype2arraycopy(T_LONG, NULL, NULL,
1203                           true, copyfunc_name, true);
1204 
1205     const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
1206     const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1207 
1208     Node* call = make_leaf_call(ctrl, mem, call_type, copyfunc_addr, copyfunc_name, raw_adr_type, src, dest, length XTOP);
1209     transform_later(call);
1210 
1211     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1212     bs->clone_barrier_at_expansion(ac, call, _igvn);
1213 
1214     return;
1215   } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_cloneoop()) {
1216     const Type* dest_type = _igvn.type(dest);
1217     const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1218 
1219     BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
1220     if (dest_elem == T_ARRAY || (dest_elem == T_VALUETYPE && top_dest->klass()->is_obj_array_klass())) {
1221       dest_elem = T_OBJECT;
1222     }
1223 
1224     Node* mem = ac->in(TypeFunc::Memory);
1225     merge_mem = MergeMemNode::make(mem);
1226     transform_later(merge_mem);
1227 
1228     AllocateArrayNode* alloc = NULL;
1229     if (ac->is_alloc_tightly_coupled()) {
1230       alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1231       assert(alloc != NULL, "expect alloc");
1232     }
1233     assert(dest_elem != T_VALUETYPE || alloc != NULL, "unsupported");
1234     Node* dest_length = alloc != NULL ? alloc->in(AllocateNode::ALength) : NULL;
1235 
1236     const TypePtr* adr_type = NULL;
1237 
1238     if (dest_elem == T_VALUETYPE) {
1239       adr_type = adjust_parameters_for_vt(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1240     } else {
1241       adr_type = _igvn.type(dest)->is_oopptr()->add_offset(Type::OffsetBot);
1242       if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1243         adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1244       }
1245       if (ac->_src_type != ac->_dest_type) {
1246         adr_type = TypeRawPtr::BOTTOM;
1247       }
1248     }
1249     generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1250                        adr_type, dest_elem,
1251                        src, src_offset, dest, dest_offset, length,
1252                        dest_length,
1253                        true, !ac->is_copyofrange());
1254 
1255     return;
1256   }
1257 
1258   AllocateArrayNode* alloc = NULL;
1259   if (ac->is_alloc_tightly_coupled()) {
1260     alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1261     assert(alloc != NULL, "expect alloc");
1262   }
1263 
1264   assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1265 
1266   // Compile time checks.  If any of these checks cannot be verified at compile time,
1267   // we do not make a fast path for this call.  Instead, we let the call remain as it
1268   // is.  The checks we choose to mandate at compile time are:
1269   //
1270   // (1) src and dest are arrays.
1271   const Type* src_type = src->Value(&_igvn);
1272   const Type* dest_type = dest->Value(&_igvn);
1273   const TypeAryPtr* top_src = src_type->isa_aryptr();
1274   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1275 
1276   BasicType src_elem = T_CONFLICT;
1277   BasicType dest_elem = T_CONFLICT;
1278 
1279   if (top_dest != NULL && top_dest->klass() != NULL) {
1280     dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
1281   }
1282   if (top_src != NULL && top_src->klass() != NULL) {
1283     src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
1284   }
1285   if (src_elem == T_ARRAY) {
1286     src_elem = T_OBJECT;
1287   } else if (src_elem == T_VALUETYPE && top_src->klass()->is_obj_array_klass()) {
1288     if (top_src->klass_is_exact()) {
1289       src_elem = T_OBJECT;
1290     } else {
1291       assert(!top_src->klass()->as_obj_array_klass()->storage_properties().is_null_free(), "klass should be exact");
1292       src_elem = T_CONFLICT; // either flattened or not
1293     }
1294   }
1295   if (dest_elem == T_ARRAY) {
1296     dest_elem = T_OBJECT;
1297   } else if (dest_elem == T_VALUETYPE && top_dest->klass()->is_obj_array_klass()) {
1298     if (top_dest->klass_is_exact()) {
1299       dest_elem = T_OBJECT;
1300     } else {
1301       assert(!top_dest->klass()->as_obj_array_klass()->storage_properties().is_null_free(), "klass should be exact");
1302       dest_elem = T_CONFLICT; // either flattened or not
1303     }
1304   }
1305 
1306   if (ac->is_arraycopy_validated() &&
1307       dest_elem != T_CONFLICT &&
1308       src_elem == T_CONFLICT) {
1309     src_elem = dest_elem;
1310   }
1311 
1312   if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1313     // Conservatively insert a memory barrier on all memory slices.
1314     // Do not let writes into the source float below the arraycopy.
1315     {
1316       Node* mem = ac->in(TypeFunc::Memory);
1317       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1318 
1319       merge_mem = MergeMemNode::make(mem);
1320       transform_later(merge_mem);
1321     }
1322 
1323     RegionNode* slow_region = new RegionNode(1);
1324     transform_later(slow_region);
1325 
1326     generate_flattened_array_guard(&ctrl, merge_mem, dest, slow_region);
1327 
1328     // Call StubRoutines::generic_arraycopy stub.
1329     Node* mem = generate_arraycopy(ac, NULL, &ctrl, merge_mem, &io,
1330                                    TypeRawPtr::BOTTOM, T_CONFLICT,
1331                                    src, src_offset, dest, dest_offset, length,
1332                                    NULL,
1333                                    // If a  negative length guard was generated for the ArrayCopyNode,
1334                                    // the length of the array can never be negative.
1335                                    false, ac->has_negative_length_guard(),
1336                                    slow_region);
1337 
1338     return;
1339   }
1340 
1341   assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1342 
1343   // (2) src and dest arrays must have elements of the same BasicType
1344   // Figure out the size and type of the elements we will be copying.
1345   //
1346   // We have no stub to copy flattened value type arrays with oop
1347   // fields if we need to emit write barriers.
1348   //
1349   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1350   if (src_elem != dest_elem || dest_elem == T_VOID) {
1351     // The component types are not the same or are not recognized.  Punt.
1352     // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1353     {
1354       Node* mem = ac->in(TypeFunc::Memory);
1355       merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1356     }
1357 
1358     _igvn.replace_node(_memproj_fallthrough, merge_mem);
1359     _igvn.replace_node(_ioproj_fallthrough, io);
1360     _igvn.replace_node(_fallthroughcatchproj, ctrl);
1361     return;
1362   }
1363 
1364   //---------------------------------------------------------------------------
1365   // We will make a fast path for this call to arraycopy.
1366 
1367   // We have the following tests left to perform:
1368   //
1369   // (3) src and dest must not be null.
1370   // (4) src_offset must not be negative.
1371   // (5) dest_offset must not be negative.
1372   // (6) length must not be negative.
1373   // (7) src_offset + length must not exceed length of src.
1374   // (8) dest_offset + length must not exceed length of dest.
1375   // (9) each element of an oop array must be assignable
1376 
1377   Node* mem = ac->in(TypeFunc::Memory);
1378   if (dest_elem == T_VALUETYPE) {
1379     // copy modifies more than 1 slice
1380     insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1381   }
1382 
1383   merge_mem = MergeMemNode::make(mem);
1384   transform_later(merge_mem);
1385 
1386   RegionNode* slow_region = new RegionNode(1);
1387   transform_later(slow_region);
1388 
1389   if (!ac->is_arraycopy_validated()) {
1390     // (3) operands must not be null
1391     // We currently perform our null checks with the null_check routine.
1392     // This means that the null exceptions will be reported in the caller
1393     // rather than (correctly) reported inside of the native arraycopy call.
1394     // This should be corrected, given time.  We do our null check with the
1395     // stack pointer restored.
1396     // null checks done library_call.cpp
1397 
1398     // (4) src_offset must not be negative.
1399     generate_negative_guard(&ctrl, src_offset, slow_region);
1400 
1401     // (5) dest_offset must not be negative.
1402     generate_negative_guard(&ctrl, dest_offset, slow_region);
1403 
1404     // (6) length must not be negative (moved to generate_arraycopy()).
1405     // generate_negative_guard(length, slow_region);
1406 
1407     // (7) src_offset + length must not exceed length of src.
1408     Node* alen = ac->in(ArrayCopyNode::SrcLen);
1409     assert(alen != NULL, "need src len");
1410     generate_limit_guard(&ctrl,
1411                          src_offset, length,
1412                          alen,
1413                          slow_region);
1414 
1415     // (8) dest_offset + length must not exceed length of dest.
1416     alen = ac->in(ArrayCopyNode::DestLen);
1417     assert(alen != NULL, "need dest len");
1418     generate_limit_guard(&ctrl,
1419                          dest_offset, length,
1420                          alen,
1421                          slow_region);
1422 
1423     // (9) each element of an oop array must be assignable
1424     // The generate_arraycopy subroutine checks this.
1425 
1426     if (dest_elem == T_OBJECT &&
1427         ValueArrayFlatten &&
1428         top_dest->elem()->make_oopptr()->can_be_value_type()) {
1429       generate_flattened_array_guard(&ctrl, merge_mem, dest, slow_region);
1430     }
1431 
1432     if (src_elem == T_OBJECT &&
1433         ValueArrayFlatten &&
1434         top_src->elem()->make_oopptr()->can_be_value_type()) {
1435       generate_flattened_array_guard(&ctrl, merge_mem, src, slow_region);
1436     }
1437   }
1438 
1439   // This is where the memory effects are placed:
1440   const TypePtr* adr_type = NULL;
1441 
1442   Node* dest_length = alloc != NULL ? alloc->in(AllocateNode::ALength) : NULL;
1443 
1444   if (dest_elem == T_VALUETYPE) {
1445     adr_type = adjust_parameters_for_vt(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1446   } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1447     adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1448   } else {
1449     adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1450   }
1451 
1452   generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1453                      adr_type, dest_elem,
1454                      src, src_offset, dest, dest_offset, length,
1455                      dest_length,
1456                      // If a  negative length guard was generated for the ArrayCopyNode,
1457                      // the length of the array can never be negative.
1458                      false, ac->has_negative_length_guard(),
1459                      slow_region);
1460 }