1 /*
   2  * Copyright (c) 2012, 2025, 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 "ci/ciFlatArrayKlass.hpp"
  26 #include "gc/shared/barrierSet.hpp"
  27 #include "gc/shared/tlab_globals.hpp"
  28 #include "oops/objArrayKlass.hpp"
  29 #include "opto/arraycopynode.hpp"
  30 #include "opto/castnode.hpp"
  31 #include "opto/convertnode.hpp"
  32 #include "opto/graphKit.hpp"
  33 #include "opto/macro.hpp"
  34 #include "opto/runtime.hpp"
  35 #include "opto/vectornode.hpp"
  36 #include "runtime/stubRoutines.hpp"
  37 #include "utilities/align.hpp"
  38 #include "utilities/powerOfTwo.hpp"
  39 
  40 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, int alias_idx, Node* precedent) {
  41   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
  42   mb->init_req(TypeFunc::Control, *ctrl);
  43   mb->init_req(TypeFunc::Memory, *mem);
  44   transform_later(mb);
  45   *ctrl = new ProjNode(mb,TypeFunc::Control);
  46   transform_later(*ctrl);
  47   Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
  48   transform_later(mem_proj);
  49   if (alias_idx == Compile::AliasIdxBot) {
  50     *mem = mem_proj;
  51   } else {
  52     MergeMemNode* mm = (*mem)->clone()->as_MergeMem();
  53     mm->set_memory_at(alias_idx, mem_proj);
  54     transform_later(mm);
  55     *mem = mm;
  56   }
  57 }
  58 
  59 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
  60   uint shift  = exact_log2(type2aelembytes(elembt));
  61   const TypeAryPtr* array_type = _igvn.type(ary)->isa_aryptr();
  62   if (array_type != nullptr && array_type->is_aryptr()->is_flat()) {
  63     // Use T_FLAT_ELEMENT to get proper alignment with COH when fetching the array element address.
  64     elembt = T_FLAT_ELEMENT;
  65   }
  66   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
  67   Node* base =  basic_plus_adr(ary, header);
  68 #ifdef _LP64
  69   // see comment in GraphKit::array_element_address
  70   int index_max = max_jint - 1;  // array size is max_jint, index is one less
  71   const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
  72   idx = transform_later( new ConvI2LNode(idx, lidxtype) );
  73 #endif
  74   Node* scale = new LShiftXNode(idx, intcon(shift));
  75   transform_later(scale);
  76   return basic_plus_adr(ary, base, scale);
  77 }
  78 
  79 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
  80   return transform_later(new ConvI2LNode(offset));
  81 }
  82 
  83 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
  84                                        const TypeFunc* call_type, address call_addr,
  85                                        const char* call_name,
  86                                        const TypePtr* adr_type,
  87                                        Node* parm0, Node* parm1,
  88                                        Node* parm2, Node* parm3,
  89                                        Node* parm4, Node* parm5,
  90                                        Node* parm6, Node* parm7) {
  91   Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
  92   call->init_req(TypeFunc::Control, ctrl);
  93   call->init_req(TypeFunc::I_O    , top());
  94   call->init_req(TypeFunc::Memory , mem);
  95   call->init_req(TypeFunc::ReturnAdr, top());
  96   call->init_req(TypeFunc::FramePtr, top());
  97 
  98   // Hook each parm in order.  Stop looking at the first null.
  99   if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
 100   if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
 101   if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
 102   if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
 103   if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
 104   if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
 105   if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
 106   if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
 107     /* close each nested if ===> */  } } } } } } } }
 108   assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
 109 
 110   return call;
 111 }
 112 
 113 
 114 //------------------------------generate_guard---------------------------
 115 // Helper function for generating guarded fast-slow graph structures.
 116 // The given 'test', if true, guards a slow path.  If the test fails
 117 // then a fast path can be taken.  (We generally hope it fails.)
 118 // In all cases, GraphKit::control() is updated to the fast path.
 119 // The returned value represents the control for the slow path.
 120 // The return value is never 'top'; it is either a valid control
 121 // or null if it is obvious that the slow path can never be taken.
 122 // Also, if region and the slow control are not null, the slow edge
 123 // is appended to the region.
 124 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
 125   if ((*ctrl)->is_top()) {
 126     // Already short circuited.
 127     return nullptr;
 128   }
 129   // Build an if node and its projections.
 130   // If test is true we take the slow path, which we assume is uncommon.
 131   if (_igvn.type(test) == TypeInt::ZERO) {
 132     // The slow branch is never taken.  No need to build this guard.
 133     return nullptr;
 134   }
 135 
 136   IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
 137   transform_later(iff);
 138 
 139   Node* if_slow = new IfTrueNode(iff);
 140   transform_later(if_slow);
 141 
 142   if (region != nullptr) {
 143     region->add_req(if_slow);
 144   }
 145 
 146   Node* if_fast = new IfFalseNode(iff);
 147   transform_later(if_fast);
 148 
 149   *ctrl = if_fast;
 150 
 151   return if_slow;
 152 }
 153 
 154 Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
 155   return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
 156 }
 157 
 158 inline Node* PhaseMacroExpand::generate_fair_guard(Node** ctrl, Node* test, RegionNode* region) {
 159   return generate_guard(ctrl, test, region, PROB_FAIR);
 160 }
 161 
 162 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
 163   if ((*ctrl)->is_top())
 164     return;                // already stopped
 165   if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
 166     return;                // index is already adequately typed
 167   Node* cmp_lt = new CmpINode(index, intcon(0));
 168   transform_later(cmp_lt);
 169   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 170   transform_later(bol_lt);
 171   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 172 }
 173 
 174 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
 175   if ((*ctrl)->is_top())
 176     return;                // already stopped
 177   bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
 178   if (zero_offset && subseq_length->eqv_uncast(array_length))
 179     return;                // common case of whole-array copy
 180   Node* last = subseq_length;
 181   if (!zero_offset) {            // last += offset
 182     last = new AddINode(last, offset);
 183     transform_later(last);
 184   }
 185   Node* cmp_lt = new CmpUNode(array_length, last);
 186   transform_later(cmp_lt);
 187   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 188   transform_later(bol_lt);
 189   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 190 }
 191 
 192 //
 193 // Partial in-lining handling for smaller conjoint/disjoint array copies having
 194 // length(in bytes) less than ArrayOperationPartialInlineSize.
 195 //  if (length <= ArrayOperationPartialInlineSize) {
 196 //    partial_inlining_block:
 197 //      mask = Mask_Gen
 198 //      vload = LoadVectorMasked src , mask
 199 //      StoreVectorMasked dst, mask, vload
 200 //  } else {
 201 //    stub_block:
 202 //      callstub array_copy
 203 //  }
 204 //  exit_block:
 205 //    Phi = label partial_inlining_block:mem , label stub_block:mem (filled by caller)
 206 //    mem = MergeMem (Phi)
 207 //    control = stub_block
 208 //
 209 //  Exit_block and associated phi(memory) are partially initialized for partial_in-lining_block
 210 //  edges. Remaining edges for exit_block coming from stub_block are connected by the caller
 211 //  post stub nodes creation.
 212 //
 213 
 214 void PhaseMacroExpand::generate_partial_inlining_block(Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type,
 215                                                        RegionNode** exit_block, Node** result_memory, Node* length,
 216                                                        Node* src_start, Node* dst_start, BasicType type) {
 217   const TypePtr *src_adr_type = _igvn.type(src_start)->isa_ptr();
 218   Node* inline_block = nullptr;
 219   Node* stub_block = nullptr;
 220 
 221   int const_len = -1;
 222   const TypeInt* lty = nullptr;
 223   uint shift  = exact_log2(type2aelembytes(type));
 224   if (length->Opcode() == Op_ConvI2L) {
 225     lty = _igvn.type(length->in(1))->isa_int();
 226   } else  {
 227     lty = _igvn.type(length)->isa_int();
 228   }
 229   if (lty && lty->is_con()) {
 230     const_len = lty->get_con() << shift;
 231   }
 232 
 233   // Return if copy length is greater than partial inline size limit or
 234   // target does not supports masked load/stores.
 235   int lane_count = ArrayCopyNode::get_partial_inline_vector_lane_count(type, const_len);
 236   if ( const_len > ArrayOperationPartialInlineSize ||
 237       !Matcher::match_rule_supported_vector(Op_LoadVectorMasked, lane_count, type)  ||
 238       !Matcher::match_rule_supported_vector(Op_StoreVectorMasked, lane_count, type) ||
 239       !Matcher::match_rule_supported_vector(Op_VectorMaskGen, lane_count, type)) {
 240     return;
 241   }
 242 
 243   int inline_limit = ArrayOperationPartialInlineSize / type2aelembytes(type);
 244   Node* casted_length = new CastLLNode(*ctrl, length, TypeLong::make(0, inline_limit, Type::WidenMin));
 245   transform_later(casted_length);
 246   Node* copy_bytes = new LShiftXNode(length, intcon(shift));
 247   transform_later(copy_bytes);
 248 
 249   Node* cmp_le = new CmpULNode(copy_bytes, longcon(ArrayOperationPartialInlineSize));
 250   transform_later(cmp_le);
 251   Node* bol_le = new BoolNode(cmp_le, BoolTest::le);
 252   transform_later(bol_le);
 253   inline_block  = generate_guard(ctrl, bol_le, nullptr, PROB_FAIR);
 254   stub_block = *ctrl;
 255 
 256   Node* mask_gen = VectorMaskGenNode::make(casted_length, type);
 257   transform_later(mask_gen);
 258 
 259   unsigned vec_size = lane_count *  type2aelembytes(type);
 260   if (C->max_vector_size() < vec_size) {
 261     C->set_max_vector_size(vec_size);
 262   }
 263 
 264   const TypeVect * vt = TypeVect::make(type, lane_count);
 265   Node* mm = (*mem)->memory_at(C->get_alias_index(src_adr_type));
 266   Node* masked_load = new LoadVectorMaskedNode(inline_block, mm, src_start,
 267                                                src_adr_type, vt, mask_gen);
 268   transform_later(masked_load);
 269 
 270   mm = (*mem)->memory_at(C->get_alias_index(adr_type));
 271   Node* masked_store = new StoreVectorMaskedNode(inline_block, mm, dst_start,
 272                                                  masked_load, adr_type, mask_gen);
 273   transform_later(masked_store);
 274 
 275   // Convergence region for inline_block and stub_block.
 276   *exit_block = new RegionNode(3);
 277   transform_later(*exit_block);
 278   (*exit_block)->init_req(1, inline_block);
 279   *result_memory = new PhiNode(*exit_block, Type::MEMORY, adr_type);
 280   transform_later(*result_memory);
 281   (*result_memory)->init_req(1, masked_store);
 282 
 283   *ctrl = stub_block;
 284 }
 285 
 286 
 287 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
 288   if ((*ctrl)->is_top())  return nullptr;
 289 
 290   if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
 291     return nullptr;                // index is already adequately typed
 292   Node* cmp_le = new CmpINode(index, intcon(0));
 293   transform_later(cmp_le);
 294   BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
 295   Node* bol_le = new BoolNode(cmp_le, le_or_eq);
 296   transform_later(bol_le);
 297   Node* is_notp = generate_guard(ctrl, bol_le, nullptr, PROB_MIN);
 298 
 299   return is_notp;
 300 }
 301 
 302 Node* PhaseMacroExpand::mark_word_test(Node** ctrl, Node* obj, MergeMemNode* mem, uintptr_t mask_val, RegionNode* region) {
 303   // Load markword and check if obj is locked
 304   Node* mark = make_load(nullptr, mem->memory_at(Compile::AliasIdxRaw), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
 305   Node* locked_bit = MakeConX(markWord::unlocked_value);
 306   locked_bit = transform_later(new AndXNode(locked_bit, mark));
 307   Node* cmp = transform_later(new CmpXNode(locked_bit, MakeConX(0)));
 308   Node* is_unlocked = transform_later(new BoolNode(cmp, BoolTest::ne));
 309   IfNode* iff = transform_later(new IfNode(*ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
 310   Node* locked_region = transform_later(new RegionNode(3));
 311   Node* mark_phi = transform_later(new PhiNode(locked_region, TypeX_X));
 312 
 313   // Unlocked: Use bits from mark word
 314   locked_region->init_req(1, transform_later(new IfTrueNode(iff)));
 315   mark_phi->init_req(1, mark);
 316 
 317   // Locked: Load prototype header from klass
 318   *ctrl = transform_later(new IfFalseNode(iff));
 319   // Make loads control dependent to make sure they are only executed if array is locked
 320   Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
 321   Node* klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
 322   Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
 323   Node* proto = transform_later(LoadNode::make(_igvn, *ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
 324 
 325   locked_region->init_req(2, *ctrl);
 326   mark_phi->init_req(2, proto);
 327   *ctrl = locked_region;
 328 
 329   // Now check if mark word bits are set
 330   Node* mask = MakeConX(mask_val);
 331   Node* masked = transform_later(new AndXNode(mark_phi, mask));
 332   cmp = transform_later(new CmpXNode(masked, mask));
 333   Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
 334   return generate_fair_guard(ctrl, bol, region);
 335 }
 336 
 337 Node* PhaseMacroExpand::generate_flat_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
 338   return mark_word_test(ctrl, array, mem, markWord::flat_array_bit_in_place, region);
 339 }
 340 
 341 Node* PhaseMacroExpand::generate_null_free_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
 342   return mark_word_test(ctrl, array, mem, markWord::null_free_array_bit_in_place, region);
 343 }
 344 
 345 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
 346   transform_later(call);
 347 
 348   *ctrl = new ProjNode(call,TypeFunc::Control);
 349   transform_later(*ctrl);
 350   Node* newmem = new ProjNode(call, TypeFunc::Memory);
 351   transform_later(newmem);
 352 
 353   uint alias_idx = C->get_alias_index(adr_type);
 354   if (alias_idx != Compile::AliasIdxBot) {
 355     *mem = MergeMemNode::make(*mem);
 356     (*mem)->set_memory_at(alias_idx, newmem);
 357   } else {
 358     *mem = MergeMemNode::make(newmem);
 359   }
 360   transform_later(*mem);
 361 }
 362 
 363 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
 364                                               Node* src_offset,
 365                                               Node* dest_offset,
 366                                               bool disjoint_bases,
 367                                               const char* &name,
 368                                               bool dest_uninitialized) {
 369   const TypeInt* src_offset_inttype  = _igvn.find_int_type(src_offset);
 370   const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);
 371 
 372   bool aligned = false;
 373   bool disjoint = disjoint_bases;
 374 
 375   // if the offsets are the same, we can treat the memory regions as
 376   // disjoint, because either the memory regions are in different arrays,
 377   // or they are identical (which we can treat as disjoint.)  We can also
 378   // treat a copy with a destination index  less that the source index
 379   // as disjoint since a low->high copy will work correctly in this case.
 380   if (src_offset_inttype != nullptr && src_offset_inttype->is_con() &&
 381       dest_offset_inttype != nullptr && dest_offset_inttype->is_con()) {
 382     // both indices are constants
 383     int s_offs = src_offset_inttype->get_con();
 384     int d_offs = dest_offset_inttype->get_con();
 385     int element_size = type2aelembytes(t);
 386     aligned = ((arrayOopDesc::base_offset_in_bytes(t) + (uint)s_offs * element_size) % HeapWordSize == 0) &&
 387               ((arrayOopDesc::base_offset_in_bytes(t) + (uint)d_offs * element_size) % HeapWordSize == 0);
 388     if (s_offs >= d_offs)  disjoint = true;
 389   } else if (src_offset == dest_offset && src_offset != nullptr) {
 390     // This can occur if the offsets are identical non-constants.
 391     disjoint = true;
 392   }
 393 
 394   return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
 395 }
 396 
 397 #define XTOP LP64_ONLY(COMMA top())
 398 
 399 // Generate an optimized call to arraycopy.
 400 // Caller must guard against non-arrays.
 401 // Caller must determine a common array basic-type for both arrays.
 402 // Caller must validate offsets against array bounds.
 403 // The slow_region has already collected guard failure paths
 404 // (such as out of bounds length or non-conformable array types).
 405 // The generated code has this shape, in general:
 406 //
 407 //     if (length == 0)  return   // via zero_path
 408 //     slowval = -1
 409 //     if (types unknown) {
 410 //       slowval = call generic copy loop
 411 //       if (slowval == 0)  return  // via checked_path
 412 //     } else if (indexes in bounds) {
 413 //       if ((is object array) && !(array type check)) {
 414 //         slowval = call checked copy loop
 415 //         if (slowval == 0)  return  // via checked_path
 416 //       } else {
 417 //         call bulk copy loop
 418 //         return  // via fast_path
 419 //       }
 420 //     }
 421 //     // adjust params for remaining work:
 422 //     if (slowval != -1) {
 423 //       n = -1^slowval; src_offset += n; dest_offset += n; length -= n
 424 //     }
 425 //   slow_region:
 426 //     call slow arraycopy(src, src_offset, dest, dest_offset, length)
 427 //     return  // via slow_call_path
 428 //
 429 // This routine is used from several intrinsics:  System.arraycopy,
 430 // Object.clone (the array subcase), and Arrays.copyOf[Range].
 431 //
 432 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
 433                                            Node** ctrl, MergeMemNode* mem, Node** io,
 434                                            const TypePtr* adr_type,
 435                                            BasicType basic_elem_type,
 436                                            Node* src,  Node* src_offset,
 437                                            Node* dest, Node* dest_offset,
 438                                            Node* copy_length,
 439                                            Node* dest_length,
 440                                            bool disjoint_bases,
 441                                            bool length_never_negative,
 442                                            RegionNode* slow_region) {
 443   if (slow_region == nullptr) {
 444     slow_region = new RegionNode(1);
 445     transform_later(slow_region);
 446   }
 447 
 448   Node* original_dest = dest;
 449   bool  dest_needs_zeroing   = false;
 450   bool  acopy_to_uninitialized = false;
 451   Node* init_value = nullptr;
 452   Node* raw_init_value = nullptr;
 453 
 454   // See if this is the initialization of a newly-allocated array.
 455   // If so, we will take responsibility here for initializing it to zero.
 456   // (Note:  Because tightly_coupled_allocation performs checks on the
 457   // out-edges of the dest, we need to avoid making derived pointers
 458   // from it until we have checked its uses.)
 459   if (ReduceBulkZeroing
 460       && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
 461       && basic_elem_type != T_CONFLICT // avoid corner case
 462       && !src->eqv_uncast(dest)
 463       && alloc != nullptr
 464       && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
 465     assert(ac->is_alloc_tightly_coupled(), "sanity");
 466     // acopy to uninitialized tightly coupled allocations
 467     // needs zeroing outside the copy range
 468     // and the acopy itself will be to uninitialized memory
 469     acopy_to_uninitialized = true;
 470     if (alloc->maybe_set_complete(&_igvn)) {
 471       // "You break it, you buy it."
 472       InitializeNode* init = alloc->initialization();
 473       assert(init->is_complete(), "we just did this");
 474       init->set_complete_with_arraycopy();
 475       assert(dest->is_CheckCastPP(), "sanity");
 476       assert(dest->in(0)->in(0) == init, "dest pinned");
 477       adr_type = TypeRawPtr::BOTTOM;  // all initializations are into raw memory
 478       // From this point on, every exit path is responsible for
 479       // initializing any non-copied parts of the object to zero.
 480       // Also, if this flag is set we make sure that arraycopy interacts properly
 481       // with G1, eliding pre-barriers. See CR 6627983.
 482       dest_needs_zeroing = true;
 483       init_value = alloc->in(AllocateNode::InitValue);
 484       raw_init_value = alloc->in(AllocateNode::RawInitValue);
 485     } else {
 486       // dest_need_zeroing = false;
 487     }
 488   } else {
 489     // No zeroing elimination needed here.
 490     alloc                  = nullptr;
 491     acopy_to_uninitialized = false;
 492     //original_dest        = dest;
 493     //dest_needs_zeroing   = false;
 494   }
 495 
 496   uint alias_idx = C->get_alias_index(adr_type);
 497 
 498   // Results are placed here:
 499   enum { fast_path        = 1,  // normal void-returning assembly stub
 500          checked_path     = 2,  // special assembly stub with cleanup
 501          slow_call_path   = 3,  // something went wrong; call the VM
 502          zero_path        = 4,  // bypass when length of copy is zero
 503          bcopy_path       = 5,  // copy primitive array by 64-bit blocks
 504          PATH_LIMIT       = 6
 505   };
 506   RegionNode* result_region = new RegionNode(PATH_LIMIT);
 507   PhiNode*    result_i_o    = new PhiNode(result_region, Type::ABIO);
 508   PhiNode*    result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
 509   assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
 510   transform_later(result_region);
 511   transform_later(result_i_o);
 512   transform_later(result_memory);
 513 
 514   // The slow_control path:
 515   Node* slow_control;
 516   Node* slow_i_o = *io;
 517   Node* slow_mem = mem->memory_at(alias_idx);
 518   DEBUG_ONLY(slow_control = (Node*) badAddress);
 519 
 520   // Checked control path:
 521   Node* checked_control = top();
 522   Node* checked_mem     = nullptr;
 523   Node* checked_i_o     = nullptr;
 524   Node* checked_value   = nullptr;
 525 
 526   if (basic_elem_type == T_CONFLICT) {
 527     assert(!dest_needs_zeroing, "");
 528     Node* cv = generate_generic_arraycopy(ctrl, &mem,
 529                                           adr_type,
 530                                           src, src_offset, dest, dest_offset,
 531                                           copy_length, acopy_to_uninitialized);
 532     if (cv == nullptr)  cv = intcon(-1);  // failure (no stub available)
 533     checked_control = *ctrl;
 534     checked_i_o     = *io;
 535     checked_mem     = mem->memory_at(alias_idx);
 536     checked_value   = cv;
 537     *ctrl = top();
 538   }
 539 
 540   Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
 541   if (not_pos != nullptr) {
 542     Node* local_ctrl = not_pos, *local_io = *io;
 543     MergeMemNode* local_mem = MergeMemNode::make(mem);
 544     transform_later(local_mem);
 545 
 546     // (6) length must not be negative.
 547     if (!length_never_negative) {
 548       generate_negative_guard(&local_ctrl, copy_length, slow_region);
 549     }
 550 
 551     // copy_length is 0.
 552     if (dest_needs_zeroing) {
 553       assert(!local_ctrl->is_top(), "no ctrl?");
 554       if (copy_length->eqv_uncast(dest_length)
 555           || _igvn.find_int_con(dest_length, 1) <= 0) {
 556         // There is no zeroing to do. No need for a secondary raw memory barrier.
 557       } else {
 558         // Clear the whole thing since there are no source elements to copy.
 559         generate_clear_array(local_ctrl, local_mem,
 560                              adr_type, dest,
 561                              init_value, raw_init_value,
 562                              basic_elem_type,
 563                              intcon(0), nullptr,
 564                              alloc->in(AllocateNode::AllocSize));
 565         // Use a secondary InitializeNode as raw memory barrier.
 566         // Currently it is needed only on this path since other
 567         // paths have stub or runtime calls as raw memory barriers.
 568         MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
 569                                           Compile::AliasIdxRaw,
 570                                           top());
 571         transform_later(mb);
 572         mb->set_req(TypeFunc::Control,local_ctrl);
 573         mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
 574         local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
 575         local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
 576 
 577         InitializeNode* init = mb->as_Initialize();
 578         init->set_complete(&_igvn);  // (there is no corresponding AllocateNode)
 579       }
 580     }
 581 
 582     // Present the results of the fast call.
 583     result_region->init_req(zero_path, local_ctrl);
 584     result_i_o   ->init_req(zero_path, local_io);
 585     result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
 586   }
 587 
 588   if (!(*ctrl)->is_top() && dest_needs_zeroing) {
 589     // We have to initialize the *uncopied* part of the array to zero.
 590     // The copy destination is the slice dest[off..off+len].  The other slices
 591     // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
 592     Node* dest_size   = alloc->in(AllocateNode::AllocSize);
 593     Node* dest_tail   = transform_later( new AddINode(dest_offset, copy_length));
 594 
 595     // If there is a head section that needs zeroing, do it now.
 596     if (_igvn.find_int_con(dest_offset, -1) != 0) {
 597       generate_clear_array(*ctrl, mem,
 598                            adr_type, dest,
 599                            init_value, raw_init_value,
 600                            basic_elem_type,
 601                            intcon(0), dest_offset,
 602                            nullptr);
 603     }
 604 
 605     // Next, perform a dynamic check on the tail length.
 606     // It is often zero, and we can win big if we prove this.
 607     // There are two wins:  Avoid generating the ClearArray
 608     // with its attendant messy index arithmetic, and upgrade
 609     // the copy to a more hardware-friendly word size of 64 bits.
 610     Node* tail_ctl = nullptr;
 611     if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
 612       Node* cmp_lt   = transform_later( new CmpINode(dest_tail, dest_length) );
 613       Node* bol_lt   = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
 614       tail_ctl = generate_slow_guard(ctrl, bol_lt, nullptr);
 615       assert(tail_ctl != nullptr || !(*ctrl)->is_top(), "must be an outcome");
 616     }
 617 
 618     // At this point, let's assume there is no tail.
 619     if (!(*ctrl)->is_top() && alloc != nullptr && basic_elem_type != T_OBJECT) {
 620       // There is no tail.  Try an upgrade to a 64-bit copy.
 621       bool didit = false;
 622       {
 623         Node* local_ctrl = *ctrl, *local_io = *io;
 624         MergeMemNode* local_mem = MergeMemNode::make(mem);
 625         transform_later(local_mem);
 626 
 627         didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
 628                                          adr_type, basic_elem_type, alloc,
 629                                          src, src_offset, dest, dest_offset,
 630                                          dest_size, acopy_to_uninitialized);
 631         if (didit) {
 632           // Present the results of the block-copying fast call.
 633           result_region->init_req(bcopy_path, local_ctrl);
 634           result_i_o   ->init_req(bcopy_path, local_io);
 635           result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
 636         }
 637       }
 638       if (didit) {
 639         *ctrl = top();     // no regular fast path
 640       }
 641     }
 642 
 643     // Clear the tail, if any.
 644     if (tail_ctl != nullptr) {
 645       Node* notail_ctl = (*ctrl)->is_top() ? nullptr : *ctrl;
 646       *ctrl = tail_ctl;
 647       if (notail_ctl == nullptr) {
 648         generate_clear_array(*ctrl, mem,
 649                              adr_type, dest,
 650                              init_value, raw_init_value,
 651                              basic_elem_type,
 652                              dest_tail, nullptr,
 653                              dest_size);
 654       } else {
 655         // Make a local merge.
 656         Node* done_ctl = transform_later(new RegionNode(3));
 657         Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
 658         done_ctl->init_req(1, notail_ctl);
 659         done_mem->init_req(1, mem->memory_at(alias_idx));
 660         generate_clear_array(*ctrl, mem,
 661                              adr_type, dest,
 662                              init_value, raw_init_value,
 663                              basic_elem_type,
 664                              dest_tail, nullptr,
 665                              dest_size);
 666         done_ctl->init_req(2, *ctrl);
 667         done_mem->init_req(2, mem->memory_at(alias_idx));
 668         *ctrl = done_ctl;
 669         mem->set_memory_at(alias_idx, done_mem);
 670       }
 671     }
 672   }
 673 
 674   BasicType copy_type = basic_elem_type;
 675   assert(basic_elem_type != T_ARRAY, "caller must fix this");
 676   if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
 677     // If src and dest have compatible element types, we can copy bits.
 678     // Types S[] and D[] are compatible if D is a supertype of S.
 679     //
 680     // If they are not, we will use checked_oop_disjoint_arraycopy,
 681     // which performs a fast optimistic per-oop check, and backs off
 682     // further to JVM_ArrayCopy on the first per-oop check that fails.
 683     // (Actually, we don't move raw bits only; the GC requires card marks.)
 684 
 685     // We don't need a subtype check for validated copies and Object[].clone()
 686     bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
 687                               ac->is_copyofrange_validated() || ac->is_clone_oop_array();
 688     if (!skip_subtype_check) {
 689       // Get the klass* for both src and dest
 690       Node* src_klass  = ac->in(ArrayCopyNode::SrcKlass);
 691       Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
 692 
 693       assert(src_klass != nullptr && dest_klass != nullptr, "should have klasses");
 694 
 695       // Generate the subtype check.
 696       // This might fold up statically, or then again it might not.
 697       //
 698       // Non-static example:  Copying List<String>.elements to a new String[].
 699       // The backing store for a List<String> is always an Object[],
 700       // but its elements are always type String, if the generic types
 701       // are correct at the source level.
 702       //
 703       // Test S[] against D[], not S against D, because (probably)
 704       // the secondary supertype cache is less busy for S[] than S.
 705       // This usually only matters when D is an interface.
 706       Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, _igvn, nullptr, -1);
 707       // Plug failing path into checked_oop_disjoint_arraycopy
 708       if (not_subtype_ctrl != top()) {
 709         Node* local_ctrl = not_subtype_ctrl;
 710         MergeMemNode* local_mem = MergeMemNode::make(mem);
 711         transform_later(local_mem);
 712 
 713         // (At this point we can assume disjoint_bases, since types differ.)
 714         int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
 715         Node* p1 = basic_plus_adr(dest_klass, ek_offset);
 716         Node* n1 = LoadKlassNode::make(_igvn, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
 717         Node* dest_elem_klass = transform_later(n1);
 718         Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
 719                                                 adr_type,
 720                                                 dest_elem_klass,
 721                                                 src, src_offset, dest, dest_offset,
 722                                                 ConvI2X(copy_length), acopy_to_uninitialized);
 723         if (cv == nullptr)  cv = intcon(-1);  // failure (no stub available)
 724         checked_control = local_ctrl;
 725         checked_i_o     = *io;
 726         checked_mem     = local_mem->memory_at(alias_idx);
 727         checked_value   = cv;
 728       }
 729     }
 730     // At this point we know we do not need type checks on oop stores.
 731 
 732     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 733     if (!bs->array_copy_requires_gc_barriers(alloc != nullptr, copy_type, false, false, BarrierSetC2::Expansion)) {
 734       // If we do not need gc barriers, copy using the jint or jlong stub.
 735       copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
 736       assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
 737              "sizes agree");
 738     }
 739   }
 740 
 741   if (!(*ctrl)->is_top()) {
 742     // Generate the fast path, if possible.
 743     Node* local_ctrl = *ctrl;
 744     MergeMemNode* local_mem = MergeMemNode::make(mem);
 745     transform_later(local_mem);
 746     generate_unchecked_arraycopy(&local_ctrl, &local_mem,
 747                                  adr_type, copy_type, disjoint_bases,
 748                                  src, src_offset, dest, dest_offset,
 749                                  ConvI2X(copy_length), acopy_to_uninitialized);
 750 
 751     // Present the results of the fast call.
 752     result_region->init_req(fast_path, local_ctrl);
 753     result_i_o   ->init_req(fast_path, *io);
 754     result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
 755   }
 756 
 757   // Here are all the slow paths up to this point, in one bundle:
 758   assert(slow_region != nullptr, "allocated on entry");
 759   slow_control = slow_region;
 760   DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
 761 
 762   *ctrl = checked_control;
 763   if (!(*ctrl)->is_top()) {
 764     // Clean up after the checked call.
 765     // The returned value is either 0 or -1^K,
 766     // where K = number of partially transferred array elements.
 767     Node* cmp = new CmpINode(checked_value, intcon(0));
 768     transform_later(cmp);
 769     Node* bol = new BoolNode(cmp, BoolTest::eq);
 770     transform_later(bol);
 771     IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
 772     transform_later(iff);
 773 
 774     // If it is 0, we are done, so transfer to the end.
 775     Node* checks_done = new IfTrueNode(iff);
 776     transform_later(checks_done);
 777     result_region->init_req(checked_path, checks_done);
 778     result_i_o   ->init_req(checked_path, checked_i_o);
 779     result_memory->init_req(checked_path, checked_mem);
 780 
 781     // If it is not zero, merge into the slow call.
 782     *ctrl = new IfFalseNode(iff);
 783     transform_later(*ctrl);
 784     RegionNode* slow_reg2 = new RegionNode(3);
 785     PhiNode*    slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
 786     PhiNode*    slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
 787     transform_later(slow_reg2);
 788     transform_later(slow_i_o2);
 789     transform_later(slow_mem2);
 790     slow_reg2  ->init_req(1, slow_control);
 791     slow_i_o2  ->init_req(1, slow_i_o);
 792     slow_mem2  ->init_req(1, slow_mem);
 793     slow_reg2  ->init_req(2, *ctrl);
 794     slow_i_o2  ->init_req(2, checked_i_o);
 795     slow_mem2  ->init_req(2, checked_mem);
 796 
 797     slow_control = slow_reg2;
 798     slow_i_o     = slow_i_o2;
 799     slow_mem     = slow_mem2;
 800 
 801     if (alloc != nullptr) {
 802       // We'll restart from the very beginning, after zeroing the whole thing.
 803       // This can cause double writes, but that's OK since dest is brand new.
 804       // So we ignore the low 31 bits of the value returned from the stub.
 805     } else {
 806       // We must continue the copy exactly where it failed, or else
 807       // another thread might see the wrong number of writes to dest.
 808       Node* checked_offset = new XorINode(checked_value, intcon(-1));
 809       Node* slow_offset    = new PhiNode(slow_reg2, TypeInt::INT);
 810       transform_later(checked_offset);
 811       transform_later(slow_offset);
 812       slow_offset->init_req(1, intcon(0));
 813       slow_offset->init_req(2, checked_offset);
 814 
 815       // Adjust the arguments by the conditionally incoming offset.
 816       Node* src_off_plus  = new AddINode(src_offset,  slow_offset);
 817       transform_later(src_off_plus);
 818       Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
 819       transform_later(dest_off_plus);
 820       Node* length_minus  = new SubINode(copy_length, slow_offset);
 821       transform_later(length_minus);
 822 
 823       // Tweak the node variables to adjust the code produced below:
 824       src_offset  = src_off_plus;
 825       dest_offset = dest_off_plus;
 826       copy_length = length_minus;
 827     }
 828   }
 829   *ctrl = slow_control;
 830   if (!(*ctrl)->is_top()) {
 831     Node* local_ctrl = *ctrl, *local_io = slow_i_o;
 832     MergeMemNode* local_mem = MergeMemNode::make(mem);
 833     transform_later(local_mem);
 834 
 835     // Generate the slow path, if needed.
 836     local_mem->set_memory_at(alias_idx, slow_mem);
 837 
 838     if (dest_needs_zeroing) {
 839       generate_clear_array(local_ctrl, local_mem,
 840                            adr_type, dest,
 841                            init_value, raw_init_value,
 842                            basic_elem_type,
 843                            intcon(0), nullptr,
 844                            alloc->in(AllocateNode::AllocSize));
 845     }
 846 
 847     local_mem = generate_slow_arraycopy(ac,
 848                                         &local_ctrl, local_mem, &local_io,
 849                                         adr_type,
 850                                         src, src_offset, dest, dest_offset,
 851                                         copy_length, /*dest_uninitialized*/false);
 852 
 853     result_region->init_req(slow_call_path, local_ctrl);
 854     result_i_o   ->init_req(slow_call_path, local_io);
 855     result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
 856   } else {
 857     ShouldNotReachHere(); // no call to generate_slow_arraycopy:
 858                           // projections were not extracted
 859   }
 860 
 861   // Remove unused edges.
 862   for (uint i = 1; i < result_region->req(); i++) {
 863     if (result_region->in(i) == nullptr) {
 864       result_region->init_req(i, top());
 865     }
 866   }
 867 
 868   // Finished; return the combined state.
 869   *ctrl = result_region;
 870   *io = result_i_o;
 871   mem->set_memory_at(alias_idx, result_memory);
 872 
 873   // mem no longer guaranteed to stay a MergeMemNode
 874   Node* out_mem = mem;
 875   DEBUG_ONLY(mem = nullptr);
 876 
 877   // The memory edges above are precise in order to model effects around
 878   // array copies accurately to allow value numbering of field loads around
 879   // arraycopy.  Such field loads, both before and after, are common in Java
 880   // collections and similar classes involving header/array data structures.
 881   //
 882   // But with low number of register or when some registers are used or killed
 883   // by arraycopy calls it causes registers spilling on stack. See 6544710.
 884   // The next memory barrier is added to avoid it. If the arraycopy can be
 885   // optimized away (which it can, sometimes) then we can manually remove
 886   // the membar also.
 887   //
 888   // Do not let reads from the cloned object float above the arraycopy.
 889   if (alloc != nullptr && !alloc->initialization()->does_not_escape()) {
 890     // Do not let stores that initialize this object be reordered with
 891     // a subsequent store that would make this object accessible by
 892     // other threads.
 893     assert(ac->_dest_type == TypeOopPtr::BOTTOM, "non escaping destination shouldn't have narrow slice");
 894     insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore, Compile::AliasIdxBot);
 895   } else {
 896     int mem_bar_alias_idx = Compile::AliasIdxBot;
 897     if (ac->_dest_type != TypeOopPtr::BOTTOM) {
 898       // The graph was transformed under the assumption the ArrayCopy node only had an effect on a narrow slice. We can't
 899       // insert a wide membar now that it's being expanded: a load that uses the input memory state of the ArrayCopy
 900       // could then become anti dependent on the membar when it was not anti dependent on the ArrayCopy leading to a
 901       // broken graph.
 902       mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
 903     }
 904     insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
 905   }
 906 
 907   assert((*ctrl)->is_Proj(), "MemBar control projection");
 908   assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
 909   (*ctrl)->in(0)->isa_MemBar()->set_trailing_expanded_array_copy();
 910 
 911   _igvn.replace_node(_callprojs->fallthrough_memproj, out_mem);
 912   if (_callprojs->fallthrough_ioproj != nullptr) {
 913     _igvn.replace_node(_callprojs->fallthrough_ioproj, *io);
 914   }
 915   _igvn.replace_node(_callprojs->fallthrough_catchproj, *ctrl);
 916 
 917 #ifdef ASSERT
 918   const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
 919   if (dest_t->is_known_instance()) {
 920     ArrayCopyNode* ac = nullptr;
 921     assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
 922     assert(ac == nullptr, "no arraycopy anymore");
 923   }
 924 #endif
 925 
 926   return out_mem;
 927 }
 928 
 929 // Helper for initialization of arrays, creating a ClearArray.
 930 // It writes zero bits in [start..end), within the body of an array object.
 931 // The memory effects are all chained onto the 'adr_type' alias category.
 932 //
 933 // Since the object is otherwise uninitialized, we are free
 934 // to put a little "slop" around the edges of the cleared area,
 935 // as long as it does not go back into the array's header,
 936 // or beyond the array end within the heap.
 937 //
 938 // The lower edge can be rounded down to the nearest jint and the
 939 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
 940 //
 941 // Arguments:
 942 //   adr_type           memory slice where writes are generated
 943 //   dest               oop of the destination array
 944 //   basic_elem_type    element type of the destination
 945 //   slice_idx          array index of first element to store
 946 //   slice_len          number of elements to store (or null)
 947 //   dest_size          total size in bytes of the array object
 948 //
 949 // Exactly one of slice_len or dest_size must be non-null.
 950 // If dest_size is non-null, zeroing extends to the end of the object.
 951 // If slice_len is non-null, the slice_idx value must be a constant.
 952 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
 953                                             const TypePtr* adr_type,
 954                                             Node* dest,
 955                                             Node* val,
 956                                             Node* raw_val,
 957                                             BasicType basic_elem_type,
 958                                             Node* slice_idx,
 959                                             Node* slice_len,
 960                                             Node* dest_size) {
 961   // one or the other but not both of slice_len and dest_size:
 962   assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
 963   if (slice_len == nullptr)  slice_len = top();
 964   if (dest_size == nullptr)  dest_size = top();
 965 
 966   uint alias_idx = C->get_alias_index(adr_type);
 967 
 968   // operate on this memory slice:
 969   Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
 970 
 971   // scaling and rounding of indexes:
 972   int scale = exact_log2(type2aelembytes(basic_elem_type));
 973   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 974   int clear_low = (-1 << scale) & (BytesPerInt  - 1);
 975   int bump_bit  = (-1 << scale) & BytesPerInt;
 976 
 977   // determine constant starts and ends
 978   const intptr_t BIG_NEG = -128;
 979   assert(BIG_NEG + 2*abase < 0, "neg enough");
 980   intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
 981   intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
 982   if (slice_len_con == 0) {
 983     return;                     // nothing to do here
 984   }
 985   intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
 986   intptr_t end_con   = _igvn.find_intptr_t_con(dest_size, -1);
 987   if (slice_idx_con >= 0 && slice_len_con >= 0) {
 988     assert(end_con < 0, "not two cons");
 989     end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
 990                        BytesPerLong);
 991   }
 992 
 993   if (start_con >= 0 && end_con >= 0) {
 994     // Constant start and end.  Simple.
 995     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
 996                                        start_con, end_con, &_igvn);
 997   } else if (start_con >= 0 && dest_size != top()) {
 998     // Constant start, pre-rounded end after the tail of the array.
 999     Node* end = dest_size;
1000     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
1001                                        start_con, end, &_igvn);
1002   } else if (start_con >= 0 && slice_len != top()) {
1003     // Constant start, non-constant end.  End needs rounding up.
1004     // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
1005     intptr_t end_base  = abase + (slice_idx_con << scale);
1006     int      end_round = (-1 << scale) & (BytesPerLong  - 1);
1007     Node*    end       = ConvI2X(slice_len);
1008     if (scale != 0)
1009       end = transform_later(new LShiftXNode(end, intcon(scale) ));
1010     end_base += end_round;
1011     end = transform_later(new AddXNode(end, MakeConX(end_base)) );
1012     end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
1013     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
1014                                        start_con, end, &_igvn);
1015   } else if (start_con < 0 && dest_size != top()) {
1016     // Non-constant start, pre-rounded end after the tail of the array.
1017     // This is almost certainly a "round-to-end" operation.
1018     Node* start = slice_idx;
1019     start = ConvI2X(start);
1020     if (scale != 0)
1021       start = transform_later(new LShiftXNode( start, intcon(scale) ));
1022     start = transform_later(new AddXNode(start, MakeConX(abase)) );
1023     if ((bump_bit | clear_low) != 0) {
1024       int to_clear = (bump_bit | clear_low);
1025       // Align up mod 8, then store a jint zero unconditionally
1026       // just before the mod-8 boundary.
1027       if (((abase + bump_bit) & ~to_clear) - bump_bit
1028           < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
1029         bump_bit = 0;
1030         assert((abase & to_clear) == 0, "array base must be long-aligned");
1031       } else {
1032         // Bump 'start' up to (or past) the next jint boundary:
1033         start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
1034         assert((abase & clear_low) == 0, "array base must be int-aligned");
1035       }
1036       // Round bumped 'start' down to jlong boundary in body of array.
1037       start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
1038       if (bump_bit != 0) {
1039         // Store a zero to the immediately preceding jint:
1040         Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
1041         Node* p1 = basic_plus_adr(dest, x1);
1042         if (val == nullptr) {
1043           assert(raw_val == nullptr, "val may not be null");
1044           mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1045         } else {
1046           assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
1047           mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
1048         }
1049         mem = transform_later(mem);
1050       }
1051     }
1052     Node* end = dest_size; // pre-rounded
1053     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
1054                                        start, end, &_igvn);
1055   } else {
1056     // Non-constant start, unrounded non-constant end.
1057     // (Nobody zeroes a random midsection of an array using this routine.)
1058     ShouldNotReachHere();       // fix caller
1059   }
1060 
1061   // Done.
1062   merge_mem->set_memory_at(alias_idx, mem);
1063 }
1064 
1065 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
1066                                                 const TypePtr* adr_type,
1067                                                 BasicType basic_elem_type,
1068                                                 AllocateNode* alloc,
1069                                                 Node* src,  Node* src_offset,
1070                                                 Node* dest, Node* dest_offset,
1071                                                 Node* dest_size, bool dest_uninitialized) {
1072   // See if there is an advantage from block transfer.
1073   int scale = exact_log2(type2aelembytes(basic_elem_type));
1074   if (scale >= LogBytesPerLong)
1075     return false;               // it is already a block transfer
1076 
1077   // Look at the alignment of the starting offsets.
1078   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1079 
1080   intptr_t src_off_con  = (intptr_t) _igvn.find_int_con(src_offset, -1);
1081   intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
1082   if (src_off_con < 0 || dest_off_con < 0) {
1083     // At present, we can only understand constants.
1084     return false;
1085   }
1086 
1087   intptr_t src_off  = abase + (src_off_con  << scale);
1088   intptr_t dest_off = abase + (dest_off_con << scale);
1089 
1090   if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1091     // Non-aligned; too bad.
1092     // One more chance:  Pick off an initial 32-bit word.
1093     // This is a common case, since abase can be odd mod 8.
1094     if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1095         ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1096       Node* sptr = basic_plus_adr(src,  src_off);
1097       Node* dptr = basic_plus_adr(dest, dest_off);
1098       const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
1099       assert(s_adr_type->isa_aryptr(), "impossible slice");
1100       uint s_alias_idx = C->get_alias_index(s_adr_type);
1101       uint d_alias_idx = C->get_alias_index(adr_type);
1102       bool is_mismatched = (basic_elem_type != T_INT);
1103       Node* sval = transform_later(
1104           LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
1105                          TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
1106                          false /*require_atomic_access*/, false /*unaligned*/, is_mismatched));
1107       Node* st = transform_later(
1108           StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
1109                           sval, T_INT, MemNode::unordered));
1110       if (is_mismatched) {
1111         st->as_Store()->set_mismatched_access();
1112       }
1113       (*mem)->set_memory_at(d_alias_idx, st);
1114       src_off += BytesPerInt;
1115       dest_off += BytesPerInt;
1116     } else {
1117       return false;
1118     }
1119   }
1120   assert(src_off % BytesPerLong == 0, "");
1121   assert(dest_off % BytesPerLong == 0, "");
1122 
1123   // Do this copy by giant steps.
1124   Node* sptr  = basic_plus_adr(src,  src_off);
1125   Node* dptr  = basic_plus_adr(dest, dest_off);
1126   Node* countx = dest_size;
1127   countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
1128   countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
1129 
1130   bool disjoint_bases = true;   // since alloc isn't null
1131   generate_unchecked_arraycopy(ctrl, mem,
1132                                adr_type, T_LONG, disjoint_bases,
1133                                sptr, nullptr, dptr, nullptr, countx, dest_uninitialized);
1134 
1135   return true;
1136 }
1137 
1138 // Helper function; generates code for the slow case.
1139 // We make a call to a runtime method which emulates the native method,
1140 // but without the native wrapper overhead.
1141 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
1142                                                         Node** ctrl, Node* mem, Node** io,
1143                                                         const TypePtr* adr_type,
1144                                                         Node* src,  Node* src_offset,
1145                                                         Node* dest, Node* dest_offset,
1146                                                         Node* copy_length, bool dest_uninitialized) {
1147   assert(!dest_uninitialized, "Invariant");
1148 
1149   const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1150   CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1151                                           "slow_arraycopy", TypePtr::BOTTOM);
1152 
1153   call->init_req(TypeFunc::Control, *ctrl);
1154   call->init_req(TypeFunc::I_O    , *io);
1155   call->init_req(TypeFunc::Memory , mem);
1156   call->init_req(TypeFunc::ReturnAdr, top());
1157   call->init_req(TypeFunc::FramePtr, top());
1158   call->init_req(TypeFunc::Parms+0, src);
1159   call->init_req(TypeFunc::Parms+1, src_offset);
1160   call->init_req(TypeFunc::Parms+2, dest);
1161   call->init_req(TypeFunc::Parms+3, dest_offset);
1162   call->init_req(TypeFunc::Parms+4, copy_length);
1163   call->copy_call_debug_info(&_igvn, ac);
1164 
1165   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1166   _igvn.replace_node(ac, call);
1167   transform_later(call);
1168 
1169   _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1170   *ctrl = _callprojs->fallthrough_catchproj->clone();
1171   transform_later(*ctrl);
1172 
1173   Node* m = _callprojs->fallthrough_memproj->clone();
1174   transform_later(m);
1175 
1176   uint alias_idx = C->get_alias_index(adr_type);
1177   MergeMemNode* out_mem;
1178   if (alias_idx != Compile::AliasIdxBot) {
1179     out_mem = MergeMemNode::make(mem);
1180     out_mem->set_memory_at(alias_idx, m);
1181   } else {
1182     out_mem = MergeMemNode::make(m);
1183   }
1184   transform_later(out_mem);
1185 
1186   // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1187   // could be nullptr. Skip clone and update nullptr fallthrough_ioproj.
1188   if (_callprojs->fallthrough_ioproj != nullptr) {
1189     *io = _callprojs->fallthrough_ioproj->clone();
1190     transform_later(*io);
1191   } else {
1192     *io = nullptr;
1193   }
1194 
1195   return out_mem;
1196 }
1197 
1198 // Helper function; generates code for cases requiring runtime checks.
1199 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1200                                                      const TypePtr* adr_type,
1201                                                      Node* dest_elem_klass,
1202                                                      Node* src,  Node* src_offset,
1203                                                      Node* dest, Node* dest_offset,
1204                                                      Node* copy_length, bool dest_uninitialized) {
1205   if ((*ctrl)->is_top())  return nullptr;
1206 
1207   address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1208   if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1209     return nullptr;
1210   }
1211 
1212   // Pick out the parameters required to perform a store-check
1213   // for the target array.  This is an optimistic check.  It will
1214   // look in each non-null element's class, at the desired klass's
1215   // super_check_offset, for the desired klass.
1216   int sco_offset = in_bytes(Klass::super_check_offset_offset());
1217   Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1218   Node* n3 = new LoadINode(nullptr, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1219   Node* check_offset = ConvI2X(transform_later(n3));
1220   Node* check_value  = dest_elem_klass;
1221 
1222   Node* src_start  = array_element_address(src,  src_offset,  T_OBJECT);
1223   Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1224 
1225   const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1226   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1227                               src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1228 
1229   finish_arraycopy_call(call, ctrl, mem, adr_type);
1230 
1231   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1232   transform_later(proj);
1233 
1234   return proj;
1235 }
1236 
1237 // Helper function; generates code for cases requiring runtime checks.
1238 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1239                                                    const TypePtr* adr_type,
1240                                                    Node* src,  Node* src_offset,
1241                                                    Node* dest, Node* dest_offset,
1242                                                    Node* copy_length, bool dest_uninitialized) {
1243   if ((*ctrl)->is_top()) return nullptr;
1244   assert(!dest_uninitialized, "Invariant");
1245 
1246   address copyfunc_addr = StubRoutines::generic_arraycopy();
1247   if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1248     return nullptr;
1249   }
1250 
1251   const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1252   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1253                               src, src_offset, dest, dest_offset, copy_length);
1254 
1255   finish_arraycopy_call(call, ctrl, mem, adr_type);
1256 
1257   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1258   transform_later(proj);
1259 
1260   return proj;
1261 }
1262 
1263 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1264 void PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1265                                                     const TypePtr* adr_type,
1266                                                     BasicType basic_elem_type,
1267                                                     bool disjoint_bases,
1268                                                     Node* src,  Node* src_offset,
1269                                                     Node* dest, Node* dest_offset,
1270                                                     Node* copy_length, bool dest_uninitialized) {
1271   if ((*ctrl)->is_top()) {
1272     return;
1273   }
1274 
1275   Node* src_start  = src;
1276   Node* dest_start = dest;
1277   if (src_offset != nullptr || dest_offset != nullptr) {
1278     src_start =  array_element_address(src, src_offset, basic_elem_type);
1279     dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1280   }
1281 
1282   // Figure out which arraycopy runtime method to call.
1283   const char* copyfunc_name = "arraycopy";
1284   address     copyfunc_addr =
1285       basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1286                           disjoint_bases, copyfunc_name, dest_uninitialized);
1287 
1288   Node* result_memory = nullptr;
1289   RegionNode* exit_block = nullptr;
1290   if (ArrayOperationPartialInlineSize > 0 && is_subword_type(basic_elem_type) &&
1291     Matcher::vector_width_in_bytes(basic_elem_type) >= 16) {
1292     generate_partial_inlining_block(ctrl, mem, adr_type, &exit_block, &result_memory,
1293                                     copy_length, src_start, dest_start, basic_elem_type);
1294   }
1295 
1296   const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1297   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1298                               src_start, dest_start, copy_length XTOP);
1299 
1300   finish_arraycopy_call(call, ctrl, mem, adr_type);
1301 
1302   // Connecting remaining edges for exit_block coming from stub_block.
1303   if (exit_block) {
1304     exit_block->init_req(2, *ctrl);
1305 
1306     // Memory edge corresponding to stub_region.
1307     result_memory->init_req(2, *mem);
1308 
1309     uint alias_idx = C->get_alias_index(adr_type);
1310     if (alias_idx != Compile::AliasIdxBot) {
1311       *mem = MergeMemNode::make(*mem);
1312       (*mem)->set_memory_at(alias_idx, result_memory);
1313     } else {
1314       *mem = MergeMemNode::make(result_memory);
1315     }
1316     transform_later(*mem);
1317     *ctrl = exit_block;
1318   }
1319 }
1320 
1321 const TypePtr* PhaseMacroExpand::adjust_for_flat_array(const TypeAryPtr* top_dest, Node*& src_offset,
1322                                                        Node*& dest_offset, Node*& length, BasicType& dest_elem,
1323                                                        Node*& dest_length) {
1324 #ifdef ASSERT
1325   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1326   bool needs_barriers = top_dest->elem()->inline_klass()->contains_oops() &&
1327     bs->array_copy_requires_gc_barriers(dest_length != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization);
1328   assert(!needs_barriers || StressReflectiveCode, "Flat arracopy would require GC barriers");
1329 #endif
1330   int elem_size = top_dest->flat_elem_size();
1331   if (elem_size >= 8) {
1332     if (elem_size > 8) {
1333       // treat as array of long but scale length, src offset and dest offset
1334       assert((elem_size % 8) == 0, "not a power of 2?");
1335       int factor = elem_size / 8;
1336       length = transform_later(new MulINode(length, intcon(factor)));
1337       src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1338       dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1339       if (dest_length != nullptr) {
1340         dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1341       }
1342       elem_size = 8;
1343     }
1344     dest_elem = T_LONG;
1345   } else if (elem_size == 4) {
1346     dest_elem = T_INT;
1347   } else if (elem_size == 2) {
1348     dest_elem = T_CHAR;
1349   } else if (elem_size == 1) {
1350     dest_elem = T_BYTE;
1351   } else {
1352     ShouldNotReachHere();
1353   }
1354   return TypeRawPtr::BOTTOM;
1355 }
1356 
1357 #undef XTOP
1358 
1359 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1360   Node* ctrl = ac->in(TypeFunc::Control);
1361   Node* io = ac->in(TypeFunc::I_O);
1362   Node* src = ac->in(ArrayCopyNode::Src);
1363   Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1364   Node* dest = ac->in(ArrayCopyNode::Dest);
1365   Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1366   Node* length = ac->in(ArrayCopyNode::Length);
1367   MergeMemNode* merge_mem = nullptr;
1368 
1369   if (ac->is_clonebasic()) {
1370     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1371     bs->clone_at_expansion(this, ac);
1372     return;
1373   } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1374     const Type* src_type = _igvn.type(src);
1375     const Type* dest_type = _igvn.type(dest);
1376     const TypeAryPtr* top_src = src_type->isa_aryptr();
1377     // Note: The destination could have type Object (i.e. non-array) when directly invoking the protected method
1378     //       Object::clone() with reflection on a declared Object that is an array at runtime. top_dest is then null.
1379     const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1380     BasicType dest_elem = T_OBJECT;
1381     if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1382       dest_elem = top_dest->elem()->array_element_basic_type();
1383     }
1384     if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1385 
1386     if (top_src != nullptr && top_src->is_flat()) {
1387       // If src is flat, dest is guaranteed to be flat as well
1388       top_dest = top_src;
1389     }
1390 
1391     AllocateArrayNode* alloc = nullptr;
1392     Node* dest_length = nullptr;
1393     if (ac->is_alloc_tightly_coupled()) {
1394       alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1395       assert(alloc != nullptr, "expect alloc");
1396       dest_length = alloc->in(AllocateNode::ALength);
1397     }
1398 
1399     Node* mem = ac->in(TypeFunc::Memory);
1400     const TypePtr* adr_type = nullptr;
1401     if (top_dest != nullptr && top_dest->is_flat()) {
1402       assert(dest_length != nullptr || StressReflectiveCode, "must be tightly coupled");
1403       // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1404       // on all slices to prevent writes into the source from floating below the arraycopy.
1405       int mem_bar_alias_idx = Compile::AliasIdxBot;
1406       if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1407         mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1408       }
1409       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1410       adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1411     } else {
1412       adr_type = dest_type->is_oopptr()->add_offset(Type::OffsetBot);
1413       if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1414         adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1415       }
1416       if (ac->_src_type != ac->_dest_type) {
1417         adr_type = TypeRawPtr::BOTTOM;
1418       }
1419     }
1420     merge_mem = MergeMemNode::make(mem);
1421     transform_later(merge_mem);
1422 
1423     generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1424                        adr_type, dest_elem,
1425                        src, src_offset, dest, dest_offset, length,
1426                        dest_length,
1427                        true, ac->has_negative_length_guard());
1428 
1429     return;
1430   }
1431 
1432   AllocateArrayNode* alloc = nullptr;
1433   if (ac->is_alloc_tightly_coupled()) {
1434     alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1435     assert(alloc != nullptr, "expect alloc");
1436   }
1437 
1438   assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1439 
1440   // Compile time checks.  If any of these checks cannot be verified at compile time,
1441   // we do not make a fast path for this call.  Instead, we let the call remain as it
1442   // is.  The checks we choose to mandate at compile time are:
1443   //
1444   // (1) src and dest are arrays.
1445   const Type* src_type = src->Value(&_igvn);
1446   const Type* dest_type = dest->Value(&_igvn);
1447   const TypeAryPtr* top_src = src_type->isa_aryptr();
1448   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1449 
1450   BasicType src_elem = T_CONFLICT;
1451   BasicType dest_elem = T_CONFLICT;
1452 
1453   if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1454     src_elem = top_src->elem()->array_element_basic_type();
1455   }
1456   if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1457     dest_elem = top_dest->elem()->array_element_basic_type();
1458   }
1459   if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1460   if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1461 
1462   if (ac->is_arraycopy_validated() && dest_elem != T_CONFLICT && src_elem == T_CONFLICT) {
1463     src_elem = dest_elem;
1464   }
1465 
1466   if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1467     // Conservatively insert a memory barrier on all memory slices.
1468     // Do not let writes into the source float below the arraycopy.
1469     {
1470       Node* mem = ac->in(TypeFunc::Memory);
1471       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1472 
1473       merge_mem = MergeMemNode::make(mem);
1474       transform_later(merge_mem);
1475     }
1476 
1477     // Call StubRoutines::generic_arraycopy stub.
1478     generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1479                        TypeRawPtr::BOTTOM, T_CONFLICT,
1480                        src, src_offset, dest, dest_offset, length,
1481                        nullptr,
1482                        // If a  negative length guard was generated for the ArrayCopyNode,
1483                        // the length of the array can never be negative.
1484                        false, ac->has_negative_length_guard());
1485     return;
1486   }
1487 
1488   assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1489 
1490   // (2) src and dest arrays must have elements of the same BasicType
1491   // Figure out the size and type of the elements we will be copying.
1492   //
1493   // We have no stub to copy flat inline type arrays with oop
1494   // fields if we need to emit write barriers.
1495   //
1496   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1497   if (src_elem != dest_elem || top_src->is_flat() != top_dest->is_flat() || dest_elem == T_VOID ||
1498       (top_src->is_flat() && top_dest->elem()->inline_klass()->contains_oops() &&
1499        bs->array_copy_requires_gc_barriers(alloc != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization))) {
1500     // The component types are not the same or are not recognized.  Punt.
1501     // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1502     {
1503       Node* mem = ac->in(TypeFunc::Memory);
1504       merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1505     }
1506 
1507     _igvn.replace_node(_callprojs->fallthrough_memproj, merge_mem);
1508     if (_callprojs->fallthrough_ioproj != nullptr) {
1509       _igvn.replace_node(_callprojs->fallthrough_ioproj, io);
1510     }
1511     _igvn.replace_node(_callprojs->fallthrough_catchproj, ctrl);
1512     return;
1513   }
1514 
1515   //---------------------------------------------------------------------------
1516   // We will make a fast path for this call to arraycopy.
1517 
1518   // We have the following tests left to perform:
1519   //
1520   // (3) src and dest must not be null.
1521   // (4) src_offset must not be negative.
1522   // (5) dest_offset must not be negative.
1523   // (6) length must not be negative.
1524   // (7) src_offset + length must not exceed length of src.
1525   // (8) dest_offset + length must not exceed length of dest.
1526   // (9) each element of an oop array must be assignable
1527 
1528   Node* mem = ac->in(TypeFunc::Memory);
1529   if (top_dest->is_flat()) {
1530     // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1531     // on all slices to prevent writes into the source from floating below the arraycopy.
1532     int mem_bar_alias_idx = Compile::AliasIdxBot;
1533     if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1534       mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1535     }
1536     insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1537   }
1538   merge_mem = MergeMemNode::make(mem);
1539   transform_later(merge_mem);
1540 
1541   RegionNode* slow_region = new RegionNode(1);
1542   transform_later(slow_region);
1543 
1544   if (!ac->is_arraycopy_validated()) {
1545     // (3) operands must not be null
1546     // We currently perform our null checks with the null_check routine.
1547     // This means that the null exceptions will be reported in the caller
1548     // rather than (correctly) reported inside of the native arraycopy call.
1549     // This should be corrected, given time.  We do our null check with the
1550     // stack pointer restored.
1551     // null checks done library_call.cpp
1552 
1553     // (4) src_offset must not be negative.
1554     generate_negative_guard(&ctrl, src_offset, slow_region);
1555 
1556     // (5) dest_offset must not be negative.
1557     generate_negative_guard(&ctrl, dest_offset, slow_region);
1558 
1559     // (6) length must not be negative (moved to generate_arraycopy()).
1560     // generate_negative_guard(length, slow_region);
1561 
1562     // (7) src_offset + length must not exceed length of src.
1563     Node* alen = ac->in(ArrayCopyNode::SrcLen);
1564     assert(alen != nullptr, "need src len");
1565     generate_limit_guard(&ctrl,
1566                          src_offset, length,
1567                          alen,
1568                          slow_region);
1569 
1570     // (8) dest_offset + length must not exceed length of dest.
1571     alen = ac->in(ArrayCopyNode::DestLen);
1572     assert(alen != nullptr, "need dest len");
1573     generate_limit_guard(&ctrl,
1574                          dest_offset, length,
1575                          alen,
1576                          slow_region);
1577 
1578     // (9) each element of an oop array must be assignable
1579     // The generate_arraycopy subroutine checks this.
1580 
1581     // TODO 8350865 Fix below logic. Also handle atomicity.
1582     // We need to be careful here because 'adjust_for_flat_array' will adjust offsets/length etc. which then does not work anymore for the slow call to SharedRuntime::slow_arraycopy_C.
1583     if (!(top_src->is_flat() && top_dest->is_flat())) {
1584       generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1585       generate_flat_array_guard(&ctrl, dest, merge_mem, slow_region);
1586     }
1587 
1588     // Handle inline type arrays
1589     if (!top_src->is_flat()) {
1590       if (UseArrayFlattening && !top_src->is_not_flat()) {
1591         // Src might be flat and dest might not be flat. Go to the slow path if src is flat.
1592         generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1593       }
1594       if (EnableValhalla) {
1595         // No validation. The subtype check emitted at macro expansion time will not go to the slow
1596         // path but call checkcast_arraycopy which can not handle flat/null-free inline type arrays.
1597         generate_null_free_array_guard(&ctrl, dest, merge_mem, slow_region);
1598       }
1599     } else {
1600       assert(top_dest->is_flat(), "dest array must be flat");
1601     }
1602   }
1603 
1604   // This is where the memory effects are placed:
1605   const TypePtr* adr_type = nullptr;
1606   Node* dest_length = (alloc != nullptr) ? alloc->in(AllocateNode::ALength) : nullptr;
1607 
1608   if (top_src->is_flat() && top_dest->is_flat()) {
1609     adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1610   } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1611     adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1612   } else {
1613     adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1614   }
1615 
1616   generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1617                      adr_type, dest_elem,
1618                      src, src_offset, dest, dest_offset, length,
1619                      dest_length,
1620                      // If a  negative length guard was generated for the ArrayCopyNode,
1621                      // the length of the array can never be negative.
1622                      false, ac->has_negative_length_guard(),
1623                      slow_region);
1624 }