1 /*
   2  * Copyright (c) 2012, 2021, 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 "ci/ciFlatArrayKlass.hpp"
  27 #include "gc/shared/barrierSet.hpp"
  28 #include "gc/shared/tlab_globals.hpp"
  29 #include "opto/arraycopynode.hpp"
  30 #include "oops/objArrayKlass.hpp"
  31 #include "opto/convertnode.hpp"
  32 #include "opto/vectornode.hpp"
  33 #include "opto/graphKit.hpp"
  34 #include "opto/macro.hpp"
  35 #include "opto/runtime.hpp"
  36 #include "opto/castnode.hpp"
  37 #include "runtime/stubRoutines.hpp"
  38 #include "utilities/align.hpp"
  39 #include "utilities/powerOfTwo.hpp"
  40 
  41 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, Node* precedent) {
  42   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
  43   mb->init_req(TypeFunc::Control, *ctrl);
  44   mb->init_req(TypeFunc::Memory, *mem);
  45   transform_later(mb);
  46   *ctrl = new ProjNode(mb,TypeFunc::Control);
  47   transform_later(*ctrl);
  48   Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
  49   transform_later(mem_proj);
  50   *mem = mem_proj;
  51 }
  52 
  53 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
  54   uint shift  = exact_log2(type2aelembytes(elembt));
  55   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
  56   Node* base =  basic_plus_adr(ary, header);
  57 #ifdef _LP64
  58   // see comment in GraphKit::array_element_address
  59   int index_max = max_jint - 1;  // array size is max_jint, index is one less
  60   const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
  61   idx = transform_later( new ConvI2LNode(idx, lidxtype) );
  62 #endif
  63   Node* scale = new LShiftXNode(idx, intcon(shift));
  64   transform_later(scale);
  65   return basic_plus_adr(ary, base, scale);
  66 }
  67 
  68 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
  69   return transform_later(new ConvI2LNode(offset));
  70 }
  71 
  72 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
  73                                        const TypeFunc* call_type, address call_addr,
  74                                        const char* call_name,
  75                                        const TypePtr* adr_type,
  76                                        Node* parm0, Node* parm1,
  77                                        Node* parm2, Node* parm3,
  78                                        Node* parm4, Node* parm5,
  79                                        Node* parm6, Node* parm7) {
  80   Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
  81   call->init_req(TypeFunc::Control, ctrl);
  82   call->init_req(TypeFunc::I_O    , top());
  83   call->init_req(TypeFunc::Memory , mem);
  84   call->init_req(TypeFunc::ReturnAdr, top());
  85   call->init_req(TypeFunc::FramePtr, top());
  86 
  87   // Hook each parm in order.  Stop looking at the first NULL.
  88   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
  89   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
  90   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
  91   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
  92   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
  93   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
  94   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
  95   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
  96     /* close each nested if ===> */  } } } } } } } }
  97   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
  98 
  99   return call;
 100 }
 101 
 102 
 103 //------------------------------generate_guard---------------------------
 104 // Helper function for generating guarded fast-slow graph structures.
 105 // The given 'test', if true, guards a slow path.  If the test fails
 106 // then a fast path can be taken.  (We generally hope it fails.)
 107 // In all cases, GraphKit::control() is updated to the fast path.
 108 // The returned value represents the control for the slow path.
 109 // The return value is never 'top'; it is either a valid control
 110 // or NULL if it is obvious that the slow path can never be taken.
 111 // Also, if region and the slow control are not NULL, the slow edge
 112 // is appended to the region.
 113 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
 114   if ((*ctrl)->is_top()) {
 115     // Already short circuited.
 116     return NULL;
 117   }
 118   // Build an if node and its projections.
 119   // If test is true we take the slow path, which we assume is uncommon.
 120   if (_igvn.type(test) == TypeInt::ZERO) {
 121     // The slow branch is never taken.  No need to build this guard.
 122     return NULL;
 123   }
 124 
 125   IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
 126   transform_later(iff);
 127 
 128   Node* if_slow = new IfTrueNode(iff);
 129   transform_later(if_slow);
 130 
 131   if (region != NULL) {
 132     region->add_req(if_slow);
 133   }
 134 
 135   Node* if_fast = new IfFalseNode(iff);
 136   transform_later(if_fast);
 137 
 138   *ctrl = if_fast;
 139 
 140   return if_slow;
 141 }
 142 
 143 Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
 144   return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
 145 }
 146 
 147 inline Node* PhaseMacroExpand::generate_fair_guard(Node** ctrl, Node* test, RegionNode* region) {
 148   return generate_guard(ctrl, test, region, PROB_FAIR);
 149 }
 150 
 151 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
 152   if ((*ctrl)->is_top())
 153     return;                // already stopped
 154   if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
 155     return;                // index is already adequately typed
 156   Node* cmp_lt = new CmpINode(index, intcon(0));
 157   transform_later(cmp_lt);
 158   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 159   transform_later(bol_lt);
 160   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 161 }
 162 
 163 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
 164   if ((*ctrl)->is_top())
 165     return;                // already stopped
 166   bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
 167   if (zero_offset && subseq_length->eqv_uncast(array_length))
 168     return;                // common case of whole-array copy
 169   Node* last = subseq_length;
 170   if (!zero_offset) {            // last += offset
 171     last = new AddINode(last, offset);
 172     transform_later(last);
 173   }
 174   Node* cmp_lt = new CmpUNode(array_length, last);
 175   transform_later(cmp_lt);
 176   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 177   transform_later(bol_lt);
 178   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 179 }
 180 
 181 //
 182 // Partial in-lining handling for smaller conjoint/disjoint array copies having
 183 // length(in bytes) less than ArrayOperationPartialInlineSize.
 184 //  if (length <= ArrayOperationPartialInlineSize) {
 185 //    partial_inlining_block:
 186 //      mask = Mask_Gen
 187 //      vload = LoadVectorMasked src , mask
 188 //      StoreVectorMasked dst, mask, vload
 189 //  } else {
 190 //    stub_block:
 191 //      callstub array_copy
 192 //  }
 193 //  exit_block:
 194 //    Phi = label partial_inlining_block:mem , label stub_block:mem (filled by caller)
 195 //    mem = MergeMem (Phi)
 196 //    control = stub_block
 197 //
 198 //  Exit_block and associated phi(memory) are partially initialized for partial_in-lining_block
 199 //  edges. Remaining edges for exit_block coming from stub_block are connected by the caller
 200 //  post stub nodes creation.
 201 //
 202 
 203 void PhaseMacroExpand::generate_partial_inlining_block(Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type,
 204                                                        RegionNode** exit_block, Node** result_memory, Node* length,
 205                                                        Node* src_start, Node* dst_start, BasicType type) {
 206   const TypePtr *src_adr_type = _igvn.type(src_start)->isa_ptr();
 207   Node* inline_block = NULL;
 208   Node* stub_block = NULL;
 209 
 210   int const_len = -1;
 211   const TypeInt* lty = NULL;
 212   uint shift  = exact_log2(type2aelembytes(type));
 213   if (length->Opcode() == Op_ConvI2L) {
 214     lty = _igvn.type(length->in(1))->isa_int();
 215   } else  {
 216     lty = _igvn.type(length)->isa_int();
 217   }
 218   if (lty && lty->is_con()) {
 219     const_len = lty->get_con() << shift;
 220   }
 221 
 222   // Return if copy length is greater than partial inline size limit or
 223   // target does not supports masked load/stores.
 224   int lane_count = ArrayCopyNode::get_partial_inline_vector_lane_count(type, const_len);
 225   if ( const_len > ArrayOperationPartialInlineSize ||
 226       !Matcher::match_rule_supported_vector(Op_LoadVectorMasked, lane_count, type)  ||
 227       !Matcher::match_rule_supported_vector(Op_StoreVectorMasked, lane_count, type) ||
 228       !Matcher::match_rule_supported_vector(Op_VectorMaskGen, lane_count, type)) {
 229     return;
 230   }
 231 
 232   int inline_limit = ArrayOperationPartialInlineSize / type2aelembytes(type);
 233   Node* casted_length = new CastLLNode(*ctrl, length, TypeLong::make(0, inline_limit, Type::WidenMin));
 234   transform_later(casted_length);
 235   Node* copy_bytes = new LShiftXNode(length, intcon(shift));
 236   transform_later(copy_bytes);
 237 
 238   Node* cmp_le = new CmpULNode(copy_bytes, longcon(ArrayOperationPartialInlineSize));
 239   transform_later(cmp_le);
 240   Node* bol_le = new BoolNode(cmp_le, BoolTest::le);
 241   transform_later(bol_le);
 242   inline_block  = generate_guard(ctrl, bol_le, NULL, PROB_FAIR);
 243   stub_block = *ctrl;
 244 
 245   Node* mask_gen = VectorMaskGenNode::make(casted_length, type);
 246   transform_later(mask_gen);
 247 
 248   unsigned vec_size = lane_count *  type2aelembytes(type);
 249   if (C->max_vector_size() < vec_size) {
 250     C->set_max_vector_size(vec_size);
 251   }
 252 
 253   const TypeVect * vt = TypeVect::make(type, lane_count);
 254   Node* mm = (*mem)->memory_at(C->get_alias_index(src_adr_type));
 255   Node* masked_load = new LoadVectorMaskedNode(inline_block, mm, src_start,
 256                                                src_adr_type, vt, mask_gen);
 257   transform_later(masked_load);
 258 
 259   mm = (*mem)->memory_at(C->get_alias_index(adr_type));
 260   Node* masked_store = new StoreVectorMaskedNode(inline_block, mm, dst_start,
 261                                                  masked_load, adr_type, mask_gen);
 262   transform_later(masked_store);
 263 
 264   // Convergence region for inline_block and stub_block.
 265   *exit_block = new RegionNode(3);
 266   transform_later(*exit_block);
 267   (*exit_block)->init_req(1, inline_block);
 268   *result_memory = new PhiNode(*exit_block, Type::MEMORY, adr_type);
 269   transform_later(*result_memory);
 270   (*result_memory)->init_req(1, masked_store);
 271 
 272   *ctrl = stub_block;
 273 }
 274 
 275 
 276 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
 277   if ((*ctrl)->is_top())  return NULL;
 278 
 279   if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
 280     return NULL;                // index is already adequately typed
 281   Node* cmp_le = new CmpINode(index, intcon(0));
 282   transform_later(cmp_le);
 283   BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
 284   Node* bol_le = new BoolNode(cmp_le, le_or_eq);
 285   transform_later(bol_le);
 286   Node* is_notp = generate_guard(ctrl, bol_le, NULL, PROB_MIN);
 287 
 288   return is_notp;
 289 }
 290 
 291 Node* PhaseMacroExpand::array_lh_test(Node* array, jint mask) {
 292   Node* klass_adr = basic_plus_adr(array, oopDesc::klass_offset_in_bytes());
 293   Node* klass = transform_later(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
 294   Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
 295   Node* lh_val = _igvn.transform(LoadNode::make(_igvn, NULL, C->immutable_memory(), lh_addr, lh_addr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
 296   Node* masked = transform_later(new AndINode(lh_val, intcon(mask)));
 297   Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
 298   return transform_later(new BoolNode(cmp, BoolTest::ne));
 299 }
 300 
 301 Node* PhaseMacroExpand::generate_flat_array_guard(Node** ctrl, Node* array, RegionNode* region) {
 302   assert(UseFlatArray, "can never be flattened");
 303   return generate_fair_guard(ctrl, array_lh_test(array, Klass::_lh_array_tag_flat_value_bit_inplace), region);
 304 }
 305 
 306 Node* PhaseMacroExpand::generate_null_free_array_guard(Node** ctrl, Node* array, RegionNode* region) {
 307   assert(EnableValhalla, "can never be null free");
 308   return generate_fair_guard(ctrl, array_lh_test(array, Klass::_lh_null_free_array_bit_inplace), region);
 309 }
 310 
 311 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
 312   transform_later(call);
 313 
 314   *ctrl = new ProjNode(call,TypeFunc::Control);
 315   transform_later(*ctrl);
 316   Node* newmem = new ProjNode(call, TypeFunc::Memory);
 317   transform_later(newmem);
 318 
 319   uint alias_idx = C->get_alias_index(adr_type);
 320   if (alias_idx != Compile::AliasIdxBot) {
 321     *mem = MergeMemNode::make(*mem);
 322     (*mem)->set_memory_at(alias_idx, newmem);
 323   } else {
 324     *mem = MergeMemNode::make(newmem);
 325   }
 326   transform_later(*mem);
 327 }
 328 
 329 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
 330                                               Node* src_offset,
 331                                               Node* dest_offset,
 332                                               bool disjoint_bases,
 333                                               const char* &name,
 334                                               bool dest_uninitialized) {
 335   const TypeInt* src_offset_inttype  = _igvn.find_int_type(src_offset);
 336   const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);
 337 
 338   bool aligned = false;
 339   bool disjoint = disjoint_bases;
 340 
 341   // if the offsets are the same, we can treat the memory regions as
 342   // disjoint, because either the memory regions are in different arrays,
 343   // or they are identical (which we can treat as disjoint.)  We can also
 344   // treat a copy with a destination index  less that the source index
 345   // as disjoint since a low->high copy will work correctly in this case.
 346   if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
 347       dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
 348     // both indices are constants
 349     int s_offs = src_offset_inttype->get_con();
 350     int d_offs = dest_offset_inttype->get_con();
 351     int element_size = type2aelembytes(t);
 352     aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
 353               ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
 354     if (s_offs >= d_offs)  disjoint = true;
 355   } else if (src_offset == dest_offset && src_offset != NULL) {
 356     // This can occur if the offsets are identical non-constants.
 357     disjoint = true;
 358   }
 359 
 360   return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
 361 }
 362 
 363 #define XTOP LP64_ONLY(COMMA top())
 364 
 365 // Generate an optimized call to arraycopy.
 366 // Caller must guard against non-arrays.
 367 // Caller must determine a common array basic-type for both arrays.
 368 // Caller must validate offsets against array bounds.
 369 // The slow_region has already collected guard failure paths
 370 // (such as out of bounds length or non-conformable array types).
 371 // The generated code has this shape, in general:
 372 //
 373 //     if (length == 0)  return   // via zero_path
 374 //     slowval = -1
 375 //     if (types unknown) {
 376 //       slowval = call generic copy loop
 377 //       if (slowval == 0)  return  // via checked_path
 378 //     } else if (indexes in bounds) {
 379 //       if ((is object array) && !(array type check)) {
 380 //         slowval = call checked copy loop
 381 //         if (slowval == 0)  return  // via checked_path
 382 //       } else {
 383 //         call bulk copy loop
 384 //         return  // via fast_path
 385 //       }
 386 //     }
 387 //     // adjust params for remaining work:
 388 //     if (slowval != -1) {
 389 //       n = -1^slowval; src_offset += n; dest_offset += n; length -= n
 390 //     }
 391 //   slow_region:
 392 //     call slow arraycopy(src, src_offset, dest, dest_offset, length)
 393 //     return  // via slow_call_path
 394 //
 395 // This routine is used from several intrinsics:  System.arraycopy,
 396 // Object.clone (the array subcase), and Arrays.copyOf[Range].
 397 //
 398 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
 399                                            Node** ctrl, MergeMemNode* mem, Node** io,
 400                                            const TypePtr* adr_type,
 401                                            BasicType basic_elem_type,
 402                                            Node* src,  Node* src_offset,
 403                                            Node* dest, Node* dest_offset,
 404                                            Node* copy_length,
 405                                            Node* dest_length,
 406                                            bool disjoint_bases,
 407                                            bool length_never_negative,
 408                                            RegionNode* slow_region) {
 409   if (slow_region == NULL) {
 410     slow_region = new RegionNode(1);
 411     transform_later(slow_region);
 412   }
 413 
 414   Node* original_dest = dest;
 415   bool  dest_needs_zeroing   = false;
 416   bool  acopy_to_uninitialized = false;
 417   Node* default_value = NULL;
 418   Node* raw_default_value = NULL;
 419 
 420   // See if this is the initialization of a newly-allocated array.
 421   // If so, we will take responsibility here for initializing it to zero.
 422   // (Note:  Because tightly_coupled_allocation performs checks on the
 423   // out-edges of the dest, we need to avoid making derived pointers
 424   // from it until we have checked its uses.)
 425   if (ReduceBulkZeroing
 426       && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
 427       && basic_elem_type != T_CONFLICT // avoid corner case
 428       && !src->eqv_uncast(dest)
 429       && alloc != NULL
 430       && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
 431     assert(ac->is_alloc_tightly_coupled(), "sanity");
 432     // acopy to uninitialized tightly coupled allocations
 433     // needs zeroing outside the copy range
 434     // and the acopy itself will be to uninitialized memory
 435     acopy_to_uninitialized = true;
 436     if (alloc->maybe_set_complete(&_igvn)) {
 437       // "You break it, you buy it."
 438       InitializeNode* init = alloc->initialization();
 439       assert(init->is_complete(), "we just did this");
 440       init->set_complete_with_arraycopy();
 441       assert(dest->is_CheckCastPP(), "sanity");
 442       assert(dest->in(0)->in(0) == init, "dest pinned");
 443       adr_type = TypeRawPtr::BOTTOM;  // all initializations are into raw memory
 444       // From this point on, every exit path is responsible for
 445       // initializing any non-copied parts of the object to zero.
 446       // Also, if this flag is set we make sure that arraycopy interacts properly
 447       // with G1, eliding pre-barriers. See CR 6627983.
 448       dest_needs_zeroing = true;
 449       default_value = alloc->in(AllocateNode::DefaultValue);
 450       raw_default_value = alloc->in(AllocateNode::RawDefaultValue);
 451     } else {
 452       // dest_need_zeroing = false;
 453     }
 454   } else {
 455     // No zeroing elimination needed here.
 456     alloc                  = NULL;
 457     acopy_to_uninitialized = false;
 458     //original_dest        = dest;
 459     //dest_needs_zeroing   = false;
 460   }
 461 
 462   uint alias_idx = C->get_alias_index(adr_type);
 463 
 464   // Results are placed here:
 465   enum { fast_path        = 1,  // normal void-returning assembly stub
 466          checked_path     = 2,  // special assembly stub with cleanup
 467          slow_call_path   = 3,  // something went wrong; call the VM
 468          zero_path        = 4,  // bypass when length of copy is zero
 469          bcopy_path       = 5,  // copy primitive array by 64-bit blocks
 470          PATH_LIMIT       = 6
 471   };
 472   RegionNode* result_region = new RegionNode(PATH_LIMIT);
 473   PhiNode*    result_i_o    = new PhiNode(result_region, Type::ABIO);
 474   PhiNode*    result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
 475   assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
 476   transform_later(result_region);
 477   transform_later(result_i_o);
 478   transform_later(result_memory);
 479 
 480   // The slow_control path:
 481   Node* slow_control;
 482   Node* slow_i_o = *io;
 483   Node* slow_mem = mem->memory_at(alias_idx);
 484   DEBUG_ONLY(slow_control = (Node*) badAddress);
 485 
 486   // Checked control path:
 487   Node* checked_control = top();
 488   Node* checked_mem     = NULL;
 489   Node* checked_i_o     = NULL;
 490   Node* checked_value   = NULL;
 491 
 492   if (basic_elem_type == T_CONFLICT) {
 493     assert(!dest_needs_zeroing, "");
 494     Node* cv = generate_generic_arraycopy(ctrl, &mem,
 495                                           adr_type,
 496                                           src, src_offset, dest, dest_offset,
 497                                           copy_length, acopy_to_uninitialized);
 498     if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
 499     checked_control = *ctrl;
 500     checked_i_o     = *io;
 501     checked_mem     = mem->memory_at(alias_idx);
 502     checked_value   = cv;
 503     *ctrl = top();
 504   }
 505 
 506   Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
 507   if (not_pos != NULL) {
 508     Node* local_ctrl = not_pos, *local_io = *io;
 509     MergeMemNode* local_mem = MergeMemNode::make(mem);
 510     transform_later(local_mem);
 511 
 512     // (6) length must not be negative.
 513     if (!length_never_negative) {
 514       generate_negative_guard(&local_ctrl, copy_length, slow_region);
 515     }
 516 
 517     // copy_length is 0.
 518     if (dest_needs_zeroing) {
 519       assert(!local_ctrl->is_top(), "no ctrl?");
 520       if (copy_length->eqv_uncast(dest_length)
 521           || _igvn.find_int_con(dest_length, 1) <= 0) {
 522         // There is no zeroing to do. No need for a secondary raw memory barrier.
 523       } else {
 524         // Clear the whole thing since there are no source elements to copy.
 525         generate_clear_array(local_ctrl, local_mem,
 526                              adr_type, dest,
 527                              default_value, raw_default_value,
 528                              basic_elem_type,
 529                              intcon(0), NULL,
 530                              alloc->in(AllocateNode::AllocSize));
 531         // Use a secondary InitializeNode as raw memory barrier.
 532         // Currently it is needed only on this path since other
 533         // paths have stub or runtime calls as raw memory barriers.
 534         MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
 535                                           Compile::AliasIdxRaw,
 536                                           top());
 537         transform_later(mb);
 538         mb->set_req(TypeFunc::Control,local_ctrl);
 539         mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
 540         local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
 541         local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
 542 
 543         InitializeNode* init = mb->as_Initialize();
 544         init->set_complete(&_igvn);  // (there is no corresponding AllocateNode)
 545       }
 546     }
 547 
 548     // Present the results of the fast call.
 549     result_region->init_req(zero_path, local_ctrl);
 550     result_i_o   ->init_req(zero_path, local_io);
 551     result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
 552   }
 553 
 554   if (!(*ctrl)->is_top() && dest_needs_zeroing) {
 555     // We have to initialize the *uncopied* part of the array to zero.
 556     // The copy destination is the slice dest[off..off+len].  The other slices
 557     // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
 558     Node* dest_size   = alloc->in(AllocateNode::AllocSize);
 559     Node* dest_tail   = transform_later( new AddINode(dest_offset, copy_length));
 560 
 561     // If there is a head section that needs zeroing, do it now.
 562     if (_igvn.find_int_con(dest_offset, -1) != 0) {
 563       generate_clear_array(*ctrl, mem,
 564                            adr_type, dest,
 565                            default_value, raw_default_value,
 566                            basic_elem_type,
 567                            intcon(0), dest_offset,
 568                            NULL);
 569     }
 570 
 571     // Next, perform a dynamic check on the tail length.
 572     // It is often zero, and we can win big if we prove this.
 573     // There are two wins:  Avoid generating the ClearArray
 574     // with its attendant messy index arithmetic, and upgrade
 575     // the copy to a more hardware-friendly word size of 64 bits.
 576     Node* tail_ctl = NULL;
 577     if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
 578       Node* cmp_lt   = transform_later( new CmpINode(dest_tail, dest_length) );
 579       Node* bol_lt   = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
 580       tail_ctl = generate_slow_guard(ctrl, bol_lt, NULL);
 581       assert(tail_ctl != NULL || !(*ctrl)->is_top(), "must be an outcome");
 582     }
 583 
 584     // At this point, let's assume there is no tail.
 585     if (!(*ctrl)->is_top() && alloc != NULL && basic_elem_type != T_OBJECT) {
 586       // There is no tail.  Try an upgrade to a 64-bit copy.
 587       bool didit = false;
 588       {
 589         Node* local_ctrl = *ctrl, *local_io = *io;
 590         MergeMemNode* local_mem = MergeMemNode::make(mem);
 591         transform_later(local_mem);
 592 
 593         didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
 594                                          adr_type, basic_elem_type, alloc,
 595                                          src, src_offset, dest, dest_offset,
 596                                          dest_size, acopy_to_uninitialized);
 597         if (didit) {
 598           // Present the results of the block-copying fast call.
 599           result_region->init_req(bcopy_path, local_ctrl);
 600           result_i_o   ->init_req(bcopy_path, local_io);
 601           result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
 602         }
 603       }
 604       if (didit) {
 605         *ctrl = top();     // no regular fast path
 606       }
 607     }
 608 
 609     // Clear the tail, if any.
 610     if (tail_ctl != NULL) {
 611       Node* notail_ctl = (*ctrl)->is_top() ? NULL : *ctrl;
 612       *ctrl = tail_ctl;
 613       if (notail_ctl == NULL) {
 614         generate_clear_array(*ctrl, mem,
 615                              adr_type, dest,
 616                              default_value, raw_default_value,
 617                              basic_elem_type,
 618                              dest_tail, NULL,
 619                              dest_size);
 620       } else {
 621         // Make a local merge.
 622         Node* done_ctl = transform_later(new RegionNode(3));
 623         Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
 624         done_ctl->init_req(1, notail_ctl);
 625         done_mem->init_req(1, mem->memory_at(alias_idx));
 626         generate_clear_array(*ctrl, mem,
 627                              adr_type, dest,
 628                              default_value, raw_default_value,
 629                              basic_elem_type,
 630                              dest_tail, NULL,
 631                              dest_size);
 632         done_ctl->init_req(2, *ctrl);
 633         done_mem->init_req(2, mem->memory_at(alias_idx));
 634         *ctrl = done_ctl;
 635         mem->set_memory_at(alias_idx, done_mem);
 636       }
 637     }
 638   }
 639 
 640   BasicType copy_type = basic_elem_type;
 641   assert(basic_elem_type != T_ARRAY, "caller must fix this");
 642   if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
 643     // If src and dest have compatible element types, we can copy bits.
 644     // Types S[] and D[] are compatible if D is a supertype of S.
 645     //
 646     // If they are not, we will use checked_oop_disjoint_arraycopy,
 647     // which performs a fast optimistic per-oop check, and backs off
 648     // further to JVM_ArrayCopy on the first per-oop check that fails.
 649     // (Actually, we don't move raw bits only; the GC requires card marks.)
 650 
 651     // We don't need a subtype check for validated copies and Object[].clone()
 652     bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
 653                               ac->is_copyofrange_validated() || ac->is_clone_oop_array();
 654     if (!skip_subtype_check) {
 655       // Get the klass* for both src and dest
 656       Node* src_klass  = ac->in(ArrayCopyNode::SrcKlass);
 657       Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
 658 
 659       assert(src_klass != NULL && dest_klass != NULL, "should have klasses");
 660 
 661       // Generate the subtype check.
 662       // This might fold up statically, or then again it might not.
 663       //
 664       // Non-static example:  Copying List<String>.elements to a new String[].
 665       // The backing store for a List<String> is always an Object[],
 666       // but its elements are always type String, if the generic types
 667       // are correct at the source level.
 668       //
 669       // Test S[] against D[], not S against D, because (probably)
 670       // the secondary supertype cache is less busy for S[] than S.
 671       // This usually only matters when D is an interface.
 672       Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, _igvn);
 673       // Plug failing path into checked_oop_disjoint_arraycopy
 674       if (not_subtype_ctrl != top()) {
 675         Node* local_ctrl = not_subtype_ctrl;
 676         MergeMemNode* local_mem = MergeMemNode::make(mem);
 677         transform_later(local_mem);
 678 
 679         // (At this point we can assume disjoint_bases, since types differ.)
 680         int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
 681         Node* p1 = basic_plus_adr(dest_klass, ek_offset);
 682         Node* n1 = LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
 683         Node* dest_elem_klass = transform_later(n1);
 684         Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
 685                                                 adr_type,
 686                                                 dest_elem_klass,
 687                                                 src, src_offset, dest, dest_offset,
 688                                                 ConvI2X(copy_length), acopy_to_uninitialized);
 689         if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
 690         checked_control = local_ctrl;
 691         checked_i_o     = *io;
 692         checked_mem     = local_mem->memory_at(alias_idx);
 693         checked_value   = cv;
 694       }
 695     }
 696     // At this point we know we do not need type checks on oop stores.
 697 
 698     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 699     if (!bs->array_copy_requires_gc_barriers(alloc != NULL, copy_type, false, false, BarrierSetC2::Expansion)) {
 700       // If we do not need gc barriers, copy using the jint or jlong stub.
 701       copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
 702       assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
 703              "sizes agree");
 704     }
 705   }
 706 
 707   bool is_partial_array_copy = false;
 708   if (!(*ctrl)->is_top()) {
 709     // Generate the fast path, if possible.
 710     Node* local_ctrl = *ctrl;
 711     MergeMemNode* local_mem = MergeMemNode::make(mem);
 712     transform_later(local_mem);
 713     is_partial_array_copy = generate_unchecked_arraycopy(&local_ctrl, &local_mem,
 714                                                          adr_type, copy_type, disjoint_bases,
 715                                                          src, src_offset, dest, dest_offset,
 716                                                          ConvI2X(copy_length), acopy_to_uninitialized);
 717 
 718     // Present the results of the fast call.
 719     result_region->init_req(fast_path, local_ctrl);
 720     result_i_o   ->init_req(fast_path, *io);
 721     result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
 722   }
 723 
 724   // Here are all the slow paths up to this point, in one bundle:
 725   assert(slow_region != NULL, "allocated on entry");
 726   slow_control = slow_region;
 727   DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
 728 
 729   *ctrl = checked_control;
 730   if (!(*ctrl)->is_top()) {
 731     // Clean up after the checked call.
 732     // The returned value is either 0 or -1^K,
 733     // where K = number of partially transferred array elements.
 734     Node* cmp = new CmpINode(checked_value, intcon(0));
 735     transform_later(cmp);
 736     Node* bol = new BoolNode(cmp, BoolTest::eq);
 737     transform_later(bol);
 738     IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
 739     transform_later(iff);
 740 
 741     // If it is 0, we are done, so transfer to the end.
 742     Node* checks_done = new IfTrueNode(iff);
 743     transform_later(checks_done);
 744     result_region->init_req(checked_path, checks_done);
 745     result_i_o   ->init_req(checked_path, checked_i_o);
 746     result_memory->init_req(checked_path, checked_mem);
 747 
 748     // If it is not zero, merge into the slow call.
 749     *ctrl = new IfFalseNode(iff);
 750     transform_later(*ctrl);
 751     RegionNode* slow_reg2 = new RegionNode(3);
 752     PhiNode*    slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
 753     PhiNode*    slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
 754     transform_later(slow_reg2);
 755     transform_later(slow_i_o2);
 756     transform_later(slow_mem2);
 757     slow_reg2  ->init_req(1, slow_control);
 758     slow_i_o2  ->init_req(1, slow_i_o);
 759     slow_mem2  ->init_req(1, slow_mem);
 760     slow_reg2  ->init_req(2, *ctrl);
 761     slow_i_o2  ->init_req(2, checked_i_o);
 762     slow_mem2  ->init_req(2, checked_mem);
 763 
 764     slow_control = slow_reg2;
 765     slow_i_o     = slow_i_o2;
 766     slow_mem     = slow_mem2;
 767 
 768     if (alloc != NULL) {
 769       // We'll restart from the very beginning, after zeroing the whole thing.
 770       // This can cause double writes, but that's OK since dest is brand new.
 771       // So we ignore the low 31 bits of the value returned from the stub.
 772     } else {
 773       // We must continue the copy exactly where it failed, or else
 774       // another thread might see the wrong number of writes to dest.
 775       Node* checked_offset = new XorINode(checked_value, intcon(-1));
 776       Node* slow_offset    = new PhiNode(slow_reg2, TypeInt::INT);
 777       transform_later(checked_offset);
 778       transform_later(slow_offset);
 779       slow_offset->init_req(1, intcon(0));
 780       slow_offset->init_req(2, checked_offset);
 781 
 782       // Adjust the arguments by the conditionally incoming offset.
 783       Node* src_off_plus  = new AddINode(src_offset,  slow_offset);
 784       transform_later(src_off_plus);
 785       Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
 786       transform_later(dest_off_plus);
 787       Node* length_minus  = new SubINode(copy_length, slow_offset);
 788       transform_later(length_minus);
 789 
 790       // Tweak the node variables to adjust the code produced below:
 791       src_offset  = src_off_plus;
 792       dest_offset = dest_off_plus;
 793       copy_length = length_minus;
 794     }
 795   }
 796   *ctrl = slow_control;
 797   if (!(*ctrl)->is_top()) {
 798     Node* local_ctrl = *ctrl, *local_io = slow_i_o;
 799     MergeMemNode* local_mem = MergeMemNode::make(mem);
 800     transform_later(local_mem);
 801 
 802     // Generate the slow path, if needed.
 803     local_mem->set_memory_at(alias_idx, slow_mem);
 804 
 805     if (dest_needs_zeroing) {
 806       generate_clear_array(local_ctrl, local_mem,
 807                            adr_type, dest,
 808                            default_value, raw_default_value,
 809                            basic_elem_type,
 810                            intcon(0), NULL,
 811                            alloc->in(AllocateNode::AllocSize));
 812     }
 813 
 814     local_mem = generate_slow_arraycopy(ac,
 815                                         &local_ctrl, local_mem, &local_io,
 816                                         adr_type,
 817                                         src, src_offset, dest, dest_offset,
 818                                         copy_length, /*dest_uninitialized*/false);
 819 
 820     result_region->init_req(slow_call_path, local_ctrl);
 821     result_i_o   ->init_req(slow_call_path, local_io);
 822     result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
 823   } else {
 824     ShouldNotReachHere(); // no call to generate_slow_arraycopy:
 825                           // projections were not extracted
 826   }
 827 
 828   // Remove unused edges.
 829   for (uint i = 1; i < result_region->req(); i++) {
 830     if (result_region->in(i) == NULL) {
 831       result_region->init_req(i, top());
 832     }
 833   }
 834 
 835   // Finished; return the combined state.
 836   *ctrl = result_region;
 837   *io = result_i_o;
 838   mem->set_memory_at(alias_idx, result_memory);
 839 
 840   // mem no longer guaranteed to stay a MergeMemNode
 841   Node* out_mem = mem;
 842   DEBUG_ONLY(mem = NULL);
 843 
 844   // The memory edges above are precise in order to model effects around
 845   // array copies accurately to allow value numbering of field loads around
 846   // arraycopy.  Such field loads, both before and after, are common in Java
 847   // collections and similar classes involving header/array data structures.
 848   //
 849   // But with low number of register or when some registers are used or killed
 850   // by arraycopy calls it causes registers spilling on stack. See 6544710.
 851   // The next memory barrier is added to avoid it. If the arraycopy can be
 852   // optimized away (which it can, sometimes) then we can manually remove
 853   // the membar also.
 854   //
 855   // Do not let reads from the cloned object float above the arraycopy.
 856   if (alloc != NULL && !alloc->initialization()->does_not_escape()) {
 857     // Do not let stores that initialize this object be reordered with
 858     // a subsequent store that would make this object accessible by
 859     // other threads.
 860     insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore);
 861   } else {
 862     insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
 863   }
 864 
 865   if (is_partial_array_copy) {
 866     assert((*ctrl)->is_Proj(), "MemBar control projection");
 867     assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
 868     (*ctrl)->in(0)->isa_MemBar()->set_trailing_partial_array_copy();
 869   }
 870 
 871   _igvn.replace_node(_callprojs->fallthrough_memproj, out_mem);
 872   if (_callprojs->fallthrough_ioproj != NULL) {
 873     _igvn.replace_node(_callprojs->fallthrough_ioproj, *io);
 874   }
 875   _igvn.replace_node(_callprojs->fallthrough_catchproj, *ctrl);
 876 
 877 #ifdef ASSERT
 878   const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
 879   if (dest_t->is_known_instance() && !is_partial_array_copy) {
 880     ArrayCopyNode* ac = NULL;
 881     assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
 882     assert(ac == NULL, "no arraycopy anymore");
 883   }
 884 #endif
 885 
 886   return out_mem;
 887 }
 888 
 889 // Helper for initialization of arrays, creating a ClearArray.
 890 // It writes zero bits in [start..end), within the body of an array object.
 891 // The memory effects are all chained onto the 'adr_type' alias category.
 892 //
 893 // Since the object is otherwise uninitialized, we are free
 894 // to put a little "slop" around the edges of the cleared area,
 895 // as long as it does not go back into the array's header,
 896 // or beyond the array end within the heap.
 897 //
 898 // The lower edge can be rounded down to the nearest jint and the
 899 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
 900 //
 901 // Arguments:
 902 //   adr_type           memory slice where writes are generated
 903 //   dest               oop of the destination array
 904 //   basic_elem_type    element type of the destination
 905 //   slice_idx          array index of first element to store
 906 //   slice_len          number of elements to store (or NULL)
 907 //   dest_size          total size in bytes of the array object
 908 //
 909 // Exactly one of slice_len or dest_size must be non-NULL.
 910 // If dest_size is non-NULL, zeroing extends to the end of the object.
 911 // If slice_len is non-NULL, the slice_idx value must be a constant.
 912 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
 913                                             const TypePtr* adr_type,
 914                                             Node* dest,
 915                                             Node* val,
 916                                             Node* raw_val,
 917                                             BasicType basic_elem_type,
 918                                             Node* slice_idx,
 919                                             Node* slice_len,
 920                                             Node* dest_size) {
 921   // one or the other but not both of slice_len and dest_size:
 922   assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
 923   if (slice_len == NULL)  slice_len = top();
 924   if (dest_size == NULL)  dest_size = top();
 925 
 926   uint alias_idx = C->get_alias_index(adr_type);
 927 
 928   // operate on this memory slice:
 929   Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
 930 
 931   // scaling and rounding of indexes:
 932   assert(basic_elem_type != T_PRIMITIVE_OBJECT, "should have been converted to a basic type copy");
 933   int scale = exact_log2(type2aelembytes(basic_elem_type));
 934   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 935   int clear_low = (-1 << scale) & (BytesPerInt  - 1);
 936   int bump_bit  = (-1 << scale) & BytesPerInt;
 937 
 938   // determine constant starts and ends
 939   const intptr_t BIG_NEG = -128;
 940   assert(BIG_NEG + 2*abase < 0, "neg enough");
 941   intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
 942   intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
 943   if (slice_len_con == 0) {
 944     return;                     // nothing to do here
 945   }
 946   intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
 947   intptr_t end_con   = _igvn.find_intptr_t_con(dest_size, -1);
 948   if (slice_idx_con >= 0 && slice_len_con >= 0) {
 949     assert(end_con < 0, "not two cons");
 950     end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
 951                        BytesPerLong);
 952   }
 953 
 954   if (start_con >= 0 && end_con >= 0) {
 955     // Constant start and end.  Simple.
 956     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
 957                                        start_con, end_con, &_igvn);
 958   } else if (start_con >= 0 && dest_size != top()) {
 959     // Constant start, pre-rounded end after the tail of the array.
 960     Node* end = dest_size;
 961     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
 962                                        start_con, end, &_igvn);
 963   } else if (start_con >= 0 && slice_len != top()) {
 964     // Constant start, non-constant end.  End needs rounding up.
 965     // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
 966     intptr_t end_base  = abase + (slice_idx_con << scale);
 967     int      end_round = (-1 << scale) & (BytesPerLong  - 1);
 968     Node*    end       = ConvI2X(slice_len);
 969     if (scale != 0)
 970       end = transform_later(new LShiftXNode(end, intcon(scale) ));
 971     end_base += end_round;
 972     end = transform_later(new AddXNode(end, MakeConX(end_base)) );
 973     end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
 974     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
 975                                        start_con, end, &_igvn);
 976   } else if (start_con < 0 && dest_size != top()) {
 977     // Non-constant start, pre-rounded end after the tail of the array.
 978     // This is almost certainly a "round-to-end" operation.
 979     Node* start = slice_idx;
 980     start = ConvI2X(start);
 981     if (scale != 0)
 982       start = transform_later(new LShiftXNode( start, intcon(scale) ));
 983     start = transform_later(new AddXNode(start, MakeConX(abase)) );
 984     if ((bump_bit | clear_low) != 0) {
 985       int to_clear = (bump_bit | clear_low);
 986       // Align up mod 8, then store a jint zero unconditionally
 987       // just before the mod-8 boundary.
 988       if (((abase + bump_bit) & ~to_clear) - bump_bit
 989           < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
 990         bump_bit = 0;
 991         assert((abase & to_clear) == 0, "array base must be long-aligned");
 992       } else {
 993         // Bump 'start' up to (or past) the next jint boundary:
 994         start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
 995         assert((abase & clear_low) == 0, "array base must be int-aligned");
 996       }
 997       // Round bumped 'start' down to jlong boundary in body of array.
 998       start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
 999       if (bump_bit != 0) {
1000         // Store a zero to the immediately preceding jint:
1001         Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
1002         Node* p1 = basic_plus_adr(dest, x1);
1003         if (val == NULL) {
1004           assert(raw_val == NULL, "val may not be null");
1005           mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1006         } else {
1007           assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
1008           mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
1009         }
1010         mem = transform_later(mem);
1011       }
1012     }
1013     Node* end = dest_size; // pre-rounded
1014     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
1015                                        start, end, &_igvn);
1016   } else {
1017     // Non-constant start, unrounded non-constant end.
1018     // (Nobody zeroes a random midsection of an array using this routine.)
1019     ShouldNotReachHere();       // fix caller
1020   }
1021 
1022   // Done.
1023   merge_mem->set_memory_at(alias_idx, mem);
1024 }
1025 
1026 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
1027                                                 const TypePtr* adr_type,
1028                                                 BasicType basic_elem_type,
1029                                                 AllocateNode* alloc,
1030                                                 Node* src,  Node* src_offset,
1031                                                 Node* dest, Node* dest_offset,
1032                                                 Node* dest_size, bool dest_uninitialized) {
1033   // See if there is an advantage from block transfer.
1034   int scale = exact_log2(type2aelembytes(basic_elem_type));
1035   if (scale >= LogBytesPerLong)
1036     return false;               // it is already a block transfer
1037 
1038   // Look at the alignment of the starting offsets.
1039   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1040 
1041   intptr_t src_off_con  = (intptr_t) _igvn.find_int_con(src_offset, -1);
1042   intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
1043   if (src_off_con < 0 || dest_off_con < 0) {
1044     // At present, we can only understand constants.
1045     return false;
1046   }
1047 
1048   intptr_t src_off  = abase + (src_off_con  << scale);
1049   intptr_t dest_off = abase + (dest_off_con << scale);
1050 
1051   if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1052     // Non-aligned; too bad.
1053     // One more chance:  Pick off an initial 32-bit word.
1054     // This is a common case, since abase can be odd mod 8.
1055     if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1056         ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1057       Node* sptr = basic_plus_adr(src,  src_off);
1058       Node* dptr = basic_plus_adr(dest, dest_off);
1059       const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
1060       assert(s_adr_type->isa_aryptr(), "impossible slice");
1061       uint s_alias_idx = C->get_alias_index(s_adr_type);
1062       uint d_alias_idx = C->get_alias_index(adr_type);
1063       bool is_mismatched = (basic_elem_type != T_INT);
1064       Node* sval = transform_later(
1065           LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
1066                          TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
1067                          false /*require_atomic_access*/, false /*unaligned*/, is_mismatched));
1068       Node* st = transform_later(
1069           StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
1070                           sval, T_INT, MemNode::unordered));
1071       if (is_mismatched) {
1072         st->as_Store()->set_mismatched_access();
1073       }
1074       (*mem)->set_memory_at(d_alias_idx, st);
1075       src_off += BytesPerInt;
1076       dest_off += BytesPerInt;
1077     } else {
1078       return false;
1079     }
1080   }
1081   assert(src_off % BytesPerLong == 0, "");
1082   assert(dest_off % BytesPerLong == 0, "");
1083 
1084   // Do this copy by giant steps.
1085   Node* sptr  = basic_plus_adr(src,  src_off);
1086   Node* dptr  = basic_plus_adr(dest, dest_off);
1087   Node* countx = dest_size;
1088   countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
1089   countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
1090 
1091   bool disjoint_bases = true;   // since alloc != NULL
1092   generate_unchecked_arraycopy(ctrl, mem,
1093                                adr_type, T_LONG, disjoint_bases,
1094                                sptr, NULL, dptr, NULL, countx, dest_uninitialized);
1095 
1096   return true;
1097 }
1098 
1099 // Helper function; generates code for the slow case.
1100 // We make a call to a runtime method which emulates the native method,
1101 // but without the native wrapper overhead.
1102 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
1103                                                         Node** ctrl, Node* mem, Node** io,
1104                                                         const TypePtr* adr_type,
1105                                                         Node* src,  Node* src_offset,
1106                                                         Node* dest, Node* dest_offset,
1107                                                         Node* copy_length, bool dest_uninitialized) {
1108   assert(!dest_uninitialized, "Invariant");
1109 
1110   const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1111   CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1112                                           "slow_arraycopy", TypePtr::BOTTOM);
1113 
1114   call->init_req(TypeFunc::Control, *ctrl);
1115   call->init_req(TypeFunc::I_O    , *io);
1116   call->init_req(TypeFunc::Memory , mem);
1117   call->init_req(TypeFunc::ReturnAdr, top());
1118   call->init_req(TypeFunc::FramePtr, top());
1119   call->init_req(TypeFunc::Parms+0, src);
1120   call->init_req(TypeFunc::Parms+1, src_offset);
1121   call->init_req(TypeFunc::Parms+2, dest);
1122   call->init_req(TypeFunc::Parms+3, dest_offset);
1123   call->init_req(TypeFunc::Parms+4, copy_length);
1124   call->copy_call_debug_info(&_igvn, ac);
1125 
1126   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1127   _igvn.replace_node(ac, call);
1128   transform_later(call);
1129 
1130   _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1131   *ctrl = _callprojs->fallthrough_catchproj->clone();
1132   transform_later(*ctrl);
1133 
1134   Node* m = _callprojs->fallthrough_memproj->clone();
1135   transform_later(m);
1136 
1137   uint alias_idx = C->get_alias_index(adr_type);
1138   MergeMemNode* out_mem;
1139   if (alias_idx != Compile::AliasIdxBot) {
1140     out_mem = MergeMemNode::make(mem);
1141     out_mem->set_memory_at(alias_idx, m);
1142   } else {
1143     out_mem = MergeMemNode::make(m);
1144   }
1145   transform_later(out_mem);
1146 
1147   // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1148   // could be NULL. Skip clone and update NULL fallthrough_ioproj.
1149   if (_callprojs->fallthrough_ioproj != NULL) {
1150     *io = _callprojs->fallthrough_ioproj->clone();
1151     transform_later(*io);
1152   } else {
1153     *io = NULL;
1154   }
1155 
1156   return out_mem;
1157 }
1158 
1159 // Helper function; generates code for cases requiring runtime checks.
1160 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1161                                                      const TypePtr* adr_type,
1162                                                      Node* dest_elem_klass,
1163                                                      Node* src,  Node* src_offset,
1164                                                      Node* dest, Node* dest_offset,
1165                                                      Node* copy_length, bool dest_uninitialized) {
1166   if ((*ctrl)->is_top())  return NULL;
1167 
1168   address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1169   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1170     return NULL;
1171   }
1172 
1173   // Pick out the parameters required to perform a store-check
1174   // for the target array.  This is an optimistic check.  It will
1175   // look in each non-null element's class, at the desired klass's
1176   // super_check_offset, for the desired klass.
1177   int sco_offset = in_bytes(Klass::super_check_offset_offset());
1178   Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1179   Node* n3 = new LoadINode(NULL, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1180   Node* check_offset = ConvI2X(transform_later(n3));
1181   Node* check_value  = dest_elem_klass;
1182 
1183   Node* src_start  = array_element_address(src,  src_offset,  T_OBJECT);
1184   Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1185 
1186   const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1187   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1188                               src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1189 
1190   finish_arraycopy_call(call, ctrl, mem, adr_type);
1191 
1192   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1193   transform_later(proj);
1194 
1195   return proj;
1196 }
1197 
1198 // Helper function; generates code for cases requiring runtime checks.
1199 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1200                                                    const TypePtr* adr_type,
1201                                                    Node* src,  Node* src_offset,
1202                                                    Node* dest, Node* dest_offset,
1203                                                    Node* copy_length, bool dest_uninitialized) {
1204   if ((*ctrl)->is_top()) return NULL;
1205   assert(!dest_uninitialized, "Invariant");
1206 
1207   address copyfunc_addr = StubRoutines::generic_arraycopy();
1208   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1209     return NULL;
1210   }
1211 
1212   const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1213   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1214                               src, src_offset, dest, dest_offset, copy_length);
1215 
1216   finish_arraycopy_call(call, ctrl, mem, adr_type);
1217 
1218   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1219   transform_later(proj);
1220 
1221   return proj;
1222 }
1223 
1224 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1225 bool PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1226                                                     const TypePtr* adr_type,
1227                                                     BasicType basic_elem_type,
1228                                                     bool disjoint_bases,
1229                                                     Node* src,  Node* src_offset,
1230                                                     Node* dest, Node* dest_offset,
1231                                                     Node* copy_length, bool dest_uninitialized) {
1232   if ((*ctrl)->is_top()) return false;
1233 
1234   Node* src_start  = src;
1235   Node* dest_start = dest;
1236   if (src_offset != NULL || dest_offset != NULL) {
1237     src_start =  array_element_address(src, src_offset, basic_elem_type);
1238     dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1239   }
1240 
1241   // Figure out which arraycopy runtime method to call.
1242   const char* copyfunc_name = "arraycopy";
1243   address     copyfunc_addr =
1244       basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1245                           disjoint_bases, copyfunc_name, dest_uninitialized);
1246 
1247   Node* result_memory = NULL;
1248   RegionNode* exit_block = NULL;
1249   if (ArrayOperationPartialInlineSize > 0 && is_subword_type(basic_elem_type) &&
1250     Matcher::vector_width_in_bytes(basic_elem_type) >= 16) {
1251     generate_partial_inlining_block(ctrl, mem, adr_type, &exit_block, &result_memory,
1252                                     copy_length, src_start, dest_start, basic_elem_type);
1253   }
1254 
1255   const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1256   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1257                               src_start, dest_start, copy_length XTOP);
1258 
1259   finish_arraycopy_call(call, ctrl, mem, adr_type);
1260 
1261   // Connecting remaining edges for exit_block coming from stub_block.
1262   if (exit_block) {
1263     exit_block->init_req(2, *ctrl);
1264 
1265     // Memory edge corresponding to stub_region.
1266     result_memory->init_req(2, *mem);
1267 
1268     uint alias_idx = C->get_alias_index(adr_type);
1269     if (alias_idx != Compile::AliasIdxBot) {
1270       *mem = MergeMemNode::make(*mem);
1271       (*mem)->set_memory_at(alias_idx, result_memory);
1272     } else {
1273       *mem = MergeMemNode::make(result_memory);
1274     }
1275     transform_later(*mem);
1276     *ctrl = exit_block;
1277     return true;
1278   }
1279   return false;
1280 }
1281 
1282 const TypePtr* PhaseMacroExpand::adjust_for_flat_array(const TypeAryPtr* top_dest, Node*& src_offset,
1283                                                        Node*& dest_offset, Node*& length, BasicType& dest_elem,
1284                                                        Node*& dest_length) {
1285 #ifdef ASSERT
1286   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1287   bool needs_barriers = top_dest->elem()->inline_klass()->contains_oops() &&
1288     bs->array_copy_requires_gc_barriers(dest_length != NULL, T_OBJECT, false, false, BarrierSetC2::Optimization);
1289   assert(!needs_barriers || StressReflectiveCode, "Flat arracopy would require GC barriers");
1290 #endif
1291   int elem_size = top_dest->flat_elem_size();
1292   if (elem_size >= 8) {
1293     if (elem_size > 8) {
1294       // treat as array of long but scale length, src offset and dest offset
1295       assert((elem_size % 8) == 0, "not a power of 2?");
1296       int factor = elem_size / 8;
1297       length = transform_later(new MulINode(length, intcon(factor)));
1298       src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1299       dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1300       if (dest_length != NULL) {
1301         dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1302       }
1303       elem_size = 8;
1304     }
1305     dest_elem = T_LONG;
1306   } else if (elem_size == 4) {
1307     dest_elem = T_INT;
1308   } else if (elem_size == 2) {
1309     dest_elem = T_CHAR;
1310   } else if (elem_size == 1) {
1311     dest_elem = T_BYTE;
1312   } else {
1313     ShouldNotReachHere();
1314   }
1315   return TypeRawPtr::BOTTOM;
1316 }
1317 
1318 #undef XTOP
1319 
1320 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1321   Node* ctrl = ac->in(TypeFunc::Control);
1322   Node* io = ac->in(TypeFunc::I_O);
1323   Node* src = ac->in(ArrayCopyNode::Src);
1324   Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1325   Node* dest = ac->in(ArrayCopyNode::Dest);
1326   Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1327   Node* length = ac->in(ArrayCopyNode::Length);
1328   MergeMemNode* merge_mem = NULL;
1329 
1330   if (ac->is_clonebasic()) {
1331     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1332     bs->clone_at_expansion(this, ac);
1333     return;
1334   } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1335     const Type* src_type = _igvn.type(src);
1336     const Type* dest_type = _igvn.type(dest);
1337     const TypeAryPtr* top_src = src_type->isa_aryptr();
1338     const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1339     BasicType dest_elem = T_OBJECT;
1340     if (top_dest != NULL && top_dest->elem() != Type::BOTTOM) {
1341       dest_elem = top_dest->elem()->array_element_basic_type();
1342     }
1343     if (dest_elem == T_ARRAY || dest_elem == T_NARROWOOP || (dest_elem == T_PRIMITIVE_OBJECT && !top_dest->is_flat())) {
1344       dest_elem = T_OBJECT;
1345     }
1346     if (top_src != NULL && top_src->is_flat()) {
1347       // If src is flat, dest is guaranteed to be flat as well
1348       dest_elem = T_PRIMITIVE_OBJECT;
1349       top_dest = top_src;
1350     }
1351 
1352     AllocateArrayNode* alloc = NULL;
1353     Node* dest_length = NULL;
1354     if (ac->is_alloc_tightly_coupled()) {
1355       alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1356       assert(alloc != NULL, "expect alloc");
1357       dest_length = alloc->in(AllocateNode::ALength);
1358     }
1359 
1360     Node* mem = ac->in(TypeFunc::Memory);
1361     const TypePtr* adr_type = NULL;
1362     if (dest_elem == T_PRIMITIVE_OBJECT) {
1363       assert(dest_length != NULL || StressReflectiveCode, "must be tightly coupled");
1364       // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1365       // on all slices to prevent writes into the source from floating below the arraycopy.
1366       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1367       adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1368     } else {
1369       adr_type = dest_type->is_oopptr()->add_offset(Type::OffsetBot);
1370       if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1371         adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1372       }
1373       if (ac->_src_type != ac->_dest_type) {
1374         adr_type = TypeRawPtr::BOTTOM;
1375       }
1376     }
1377     merge_mem = MergeMemNode::make(mem);
1378     transform_later(merge_mem);
1379 
1380     generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1381                        adr_type, dest_elem,
1382                        src, src_offset, dest, dest_offset, length,
1383                        dest_length,
1384                        true, !ac->is_copyofrange());
1385     return;
1386   }
1387 
1388   AllocateArrayNode* alloc = NULL;
1389   if (ac->is_alloc_tightly_coupled()) {
1390     alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1391     assert(alloc != NULL, "expect alloc");
1392   }
1393 
1394   assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1395 
1396   // Compile time checks.  If any of these checks cannot be verified at compile time,
1397   // we do not make a fast path for this call.  Instead, we let the call remain as it
1398   // is.  The checks we choose to mandate at compile time are:
1399   //
1400   // (1) src and dest are arrays.
1401   const Type* src_type = src->Value(&_igvn);
1402   const Type* dest_type = dest->Value(&_igvn);
1403   const TypeAryPtr* top_src = src_type->isa_aryptr();
1404   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1405 
1406   BasicType src_elem = T_CONFLICT;
1407   BasicType dest_elem = T_CONFLICT;
1408 
1409   if (top_src != NULL && top_src->elem() != Type::BOTTOM) {
1410     src_elem = top_src->elem()->array_element_basic_type();
1411   }
1412   if (top_dest != NULL && top_dest->elem() != Type::BOTTOM) {
1413     dest_elem = top_dest->elem()->array_element_basic_type();
1414   }
1415   if (src_elem == T_ARRAY || src_elem == T_NARROWOOP || (src_elem == T_PRIMITIVE_OBJECT && !top_src->is_flat())) {
1416     src_elem = T_OBJECT;
1417   }
1418   if (dest_elem == T_ARRAY || dest_elem == T_NARROWOOP || (dest_elem == T_PRIMITIVE_OBJECT && !top_dest->is_flat())) {
1419     dest_elem = T_OBJECT;
1420   }
1421 
1422   if (ac->is_arraycopy_validated() && dest_elem != T_CONFLICT && src_elem == T_CONFLICT) {
1423     src_elem = dest_elem;
1424   }
1425 
1426   if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1427     // Conservatively insert a memory barrier on all memory slices.
1428     // Do not let writes into the source float below the arraycopy.
1429     {
1430       Node* mem = ac->in(TypeFunc::Memory);
1431       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1432 
1433       merge_mem = MergeMemNode::make(mem);
1434       transform_later(merge_mem);
1435     }
1436 
1437     // Call StubRoutines::generic_arraycopy stub.
1438     Node* mem = generate_arraycopy(ac, NULL, &ctrl, merge_mem, &io,
1439                                    TypeRawPtr::BOTTOM, T_CONFLICT,
1440                                    src, src_offset, dest, dest_offset, length,
1441                                    NULL,
1442                                    // If a  negative length guard was generated for the ArrayCopyNode,
1443                                    // the length of the array can never be negative.
1444                                    false, ac->has_negative_length_guard());
1445     return;
1446   }
1447 
1448   assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID) ||
1449          (src_elem == T_PRIMITIVE_OBJECT && StressReflectiveCode), "validated but different basic types");
1450 
1451   // (2) src and dest arrays must have elements of the same BasicType
1452   // Figure out the size and type of the elements we will be copying.
1453   //
1454   // We have no stub to copy flattened inline type arrays with oop
1455   // fields if we need to emit write barriers.
1456   //
1457   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1458   if (src_elem != dest_elem || dest_elem == T_VOID ||
1459       (dest_elem == T_PRIMITIVE_OBJECT && top_dest->elem()->inline_klass()->contains_oops() &&
1460        bs->array_copy_requires_gc_barriers(alloc != NULL, T_OBJECT, false, false, BarrierSetC2::Optimization))) {
1461     // The component types are not the same or are not recognized.  Punt.
1462     // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1463     {
1464       Node* mem = ac->in(TypeFunc::Memory);
1465       merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1466     }
1467 
1468     _igvn.replace_node(_callprojs->fallthrough_memproj, merge_mem);
1469     if (_callprojs->fallthrough_ioproj != NULL) {
1470       _igvn.replace_node(_callprojs->fallthrough_ioproj, io);
1471     }
1472     _igvn.replace_node(_callprojs->fallthrough_catchproj, ctrl);
1473     return;
1474   }
1475 
1476   //---------------------------------------------------------------------------
1477   // We will make a fast path for this call to arraycopy.
1478 
1479   // We have the following tests left to perform:
1480   //
1481   // (3) src and dest must not be null.
1482   // (4) src_offset must not be negative.
1483   // (5) dest_offset must not be negative.
1484   // (6) length must not be negative.
1485   // (7) src_offset + length must not exceed length of src.
1486   // (8) dest_offset + length must not exceed length of dest.
1487   // (9) each element of an oop array must be assignable
1488 
1489   Node* mem = ac->in(TypeFunc::Memory);
1490   if (dest_elem == T_PRIMITIVE_OBJECT) {
1491     // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1492     // on all slices to prevent writes into the source from floating below the arraycopy.
1493     insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1494   }
1495   merge_mem = MergeMemNode::make(mem);
1496   transform_later(merge_mem);
1497 
1498   RegionNode* slow_region = new RegionNode(1);
1499   transform_later(slow_region);
1500 
1501   if (!ac->is_arraycopy_validated()) {
1502     // (3) operands must not be null
1503     // We currently perform our null checks with the null_check routine.
1504     // This means that the null exceptions will be reported in the caller
1505     // rather than (correctly) reported inside of the native arraycopy call.
1506     // This should be corrected, given time.  We do our null check with the
1507     // stack pointer restored.
1508     // null checks done library_call.cpp
1509 
1510     // (4) src_offset must not be negative.
1511     generate_negative_guard(&ctrl, src_offset, slow_region);
1512 
1513     // (5) dest_offset must not be negative.
1514     generate_negative_guard(&ctrl, dest_offset, slow_region);
1515 
1516     // (6) length must not be negative (moved to generate_arraycopy()).
1517     // generate_negative_guard(length, slow_region);
1518 
1519     // (7) src_offset + length must not exceed length of src.
1520     Node* alen = ac->in(ArrayCopyNode::SrcLen);
1521     assert(alen != NULL, "need src len");
1522     generate_limit_guard(&ctrl,
1523                          src_offset, length,
1524                          alen,
1525                          slow_region);
1526 
1527     // (8) dest_offset + length must not exceed length of dest.
1528     alen = ac->in(ArrayCopyNode::DestLen);
1529     assert(alen != NULL, "need dest len");
1530     generate_limit_guard(&ctrl,
1531                          dest_offset, length,
1532                          alen,
1533                          slow_region);
1534 
1535     // (9) each element of an oop array must be assignable
1536     // The generate_arraycopy subroutine checks this.
1537 
1538     // Handle inline type arrays
1539     if (!top_src->is_flat()) {
1540       if (UseFlatArray && !top_src->is_not_flat()) {
1541         // Src might be flat and dest might not be flat. Go to the slow path if src is flat.
1542         generate_flat_array_guard(&ctrl, src, slow_region);
1543       }
1544       if (EnableValhalla) {
1545         // No validation. The subtype check emitted at macro expansion time will not go to the slow
1546         // path but call checkcast_arraycopy which can not handle flat/null-free inline type arrays.
1547         generate_null_free_array_guard(&ctrl, dest, slow_region);
1548       }
1549     } else {
1550       assert(top_dest->is_flat(), "dest array must be flat");
1551     }
1552   }
1553 
1554   // This is where the memory effects are placed:
1555   const TypePtr* adr_type = NULL;
1556   Node* dest_length = (alloc != NULL) ? alloc->in(AllocateNode::ALength) : NULL;
1557 
1558   if (dest_elem == T_PRIMITIVE_OBJECT) {
1559     adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1560   } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1561     adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1562   } else {
1563     adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1564   }
1565 
1566   generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1567                      adr_type, dest_elem,
1568                      src, src_offset, dest, dest_offset, length,
1569                      dest_length,
1570                      // If a  negative length guard was generated for the ArrayCopyNode,
1571                      // the length of the array can never be negative.
1572                      false, ac->has_negative_length_guard(),
1573                      slow_region);
1574 }