1 /*
   2  * Copyright (c) 2012, 2026, 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, bool raw_base) {
  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, raw_base);
  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(raw_base ? top() : 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   int inline_limit = ArrayOperationPartialInlineSize / type2aelembytes(type);
 218 
 219   const TypeLong* length_type = _igvn.type(length)->isa_long();
 220   if (length_type == nullptr) {
 221     assert(_igvn.type(length) == Type::TOP, "");
 222     return;
 223   }
 224 
 225   const TypeLong* inline_range = TypeLong::make(0, inline_limit, Type::WidenMin);
 226   if (length_type->join(inline_range) == Type::TOP) {
 227     // The ranges do not intersect, the inline check will surely fail
 228     return;
 229   }
 230 
 231   // Return if the target does not supports masked load/stores.
 232   int lane_count = ArrayCopyNode::get_partial_inline_vector_lane_count(type, length_type->_hi);
 233   if (!Matcher::match_rule_supported_vector(Op_LoadVectorMasked, lane_count, type)  ||
 234       !Matcher::match_rule_supported_vector(Op_StoreVectorMasked, lane_count, type) ||
 235       !Matcher::match_rule_supported_vector(Op_VectorMaskGen, lane_count, type)) {
 236     return;
 237   }
 238 
 239   Node* cmp_le = new CmpULNode(length, longcon(inline_limit));
 240   transform_later(cmp_le);
 241   Node* bol_le = new BoolNode(cmp_le, BoolTest::le);
 242   transform_later(bol_le);
 243   Node* inline_block = generate_guard(ctrl, bol_le, nullptr, PROB_FAIR);
 244   Node* stub_block = *ctrl;
 245 
 246   Node* casted_length = new CastLLNode(inline_block, length, inline_range, ConstraintCastNode::DependencyType::FloatingNarrowing);
 247   transform_later(casted_length);
 248   Node* mask_gen = VectorMaskGenNode::make(casted_length, type);
 249   transform_later(mask_gen);
 250 
 251   unsigned vec_size = lane_count * type2aelembytes(type);
 252   if (C->max_vector_size() < vec_size) {
 253     C->set_max_vector_size(vec_size);
 254   }
 255 
 256   const TypePtr* src_adr_type = _igvn.type(src_start)->isa_ptr();
 257   const TypeVect * vt = TypeVect::make(type, lane_count);
 258   Node* mm = (*mem)->memory_at(C->get_alias_index(src_adr_type));
 259   Node* masked_load = new LoadVectorMaskedNode(inline_block, mm, src_start,
 260                                                src_adr_type, vt, mask_gen);
 261   transform_later(masked_load);
 262 
 263   mm = (*mem)->memory_at(C->get_alias_index(adr_type));
 264   Node* masked_store = new StoreVectorMaskedNode(inline_block, mm, dst_start,
 265                                                  masked_load, adr_type, mask_gen);
 266   transform_later(masked_store);
 267 
 268   // Convergence region for inline_block and stub_block.
 269   *exit_block = new RegionNode(3);
 270   transform_later(*exit_block);
 271   (*exit_block)->init_req(1, inline_block);
 272   *result_memory = new PhiNode(*exit_block, Type::MEMORY, adr_type);
 273   transform_later(*result_memory);
 274   (*result_memory)->init_req(1, masked_store);
 275 
 276   *ctrl = stub_block;
 277 }
 278 
 279 
 280 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
 281   if ((*ctrl)->is_top())  return nullptr;
 282 
 283   if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
 284     return nullptr;                // index is already adequately typed
 285   Node* cmp_le = new CmpINode(index, intcon(0));
 286   transform_later(cmp_le);
 287   BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
 288   Node* bol_le = new BoolNode(cmp_le, le_or_eq);
 289   transform_later(bol_le);
 290   Node* is_notp = generate_guard(ctrl, bol_le, nullptr, PROB_MIN);
 291 
 292   return is_notp;
 293 }
 294 
 295 Node* PhaseMacroExpand::mark_word_test(Node** ctrl, Node* obj, MergeMemNode* mem, uintptr_t mask_val, RegionNode* region) {
 296   // Load markword and check if obj is locked
 297   Node* mark = make_load_raw(nullptr, mem->memory_at(Compile::AliasIdxRaw), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
 298   Node* locked_bit = MakeConX(markWord::unlocked_value);
 299   locked_bit = transform_later(new AndXNode(locked_bit, mark));
 300   Node* cmp = transform_later(new CmpXNode(locked_bit, MakeConX(0)));
 301   Node* is_unlocked = transform_later(new BoolNode(cmp, BoolTest::ne));
 302   IfNode* iff = transform_later(new IfNode(*ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
 303   Node* locked_region = transform_later(new RegionNode(3));
 304   Node* mark_phi = transform_later(new PhiNode(locked_region, TypeX_X));
 305 
 306   // Unlocked: Use bits from mark word
 307   locked_region->init_req(1, transform_later(new IfTrueNode(iff)));
 308   mark_phi->init_req(1, mark);
 309 
 310   // Locked: Load prototype header from klass
 311   *ctrl = transform_later(new IfFalseNode(iff));
 312   // Make loads control dependent to make sure they are only executed if array is locked
 313   Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
 314   Node* klass = transform_later(LoadKlassNode::make(_igvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
 315   Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
 316   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));
 317 
 318   locked_region->init_req(2, *ctrl);
 319   mark_phi->init_req(2, proto);
 320   *ctrl = locked_region;
 321 
 322   // Now check if mark word bits are set
 323   Node* mask = MakeConX(mask_val);
 324   Node* masked = transform_later(new AndXNode(mark_phi, mask));
 325   cmp = transform_later(new CmpXNode(masked, mask));
 326   Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
 327   return generate_fair_guard(ctrl, bol, region);
 328 }
 329 
 330 Node* PhaseMacroExpand::generate_flat_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
 331   return mark_word_test(ctrl, array, mem, markWord::flat_array_bit_in_place, region);
 332 }
 333 
 334 Node* PhaseMacroExpand::generate_null_free_array_guard(Node** ctrl, Node* array, MergeMemNode* mem, RegionNode* region) {
 335   return mark_word_test(ctrl, array, mem, markWord::null_free_array_bit_in_place, region);
 336 }
 337 
 338 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
 339   transform_later(call);
 340 
 341   *ctrl = new ProjNode(call,TypeFunc::Control);
 342   transform_later(*ctrl);
 343   Node* newmem = new ProjNode(call, TypeFunc::Memory);
 344   transform_later(newmem);
 345 
 346   uint alias_idx = C->get_alias_index(adr_type);
 347   if (alias_idx != Compile::AliasIdxBot) {
 348     *mem = MergeMemNode::make(*mem);
 349     (*mem)->set_memory_at(alias_idx, newmem);
 350   } else {
 351     *mem = MergeMemNode::make(newmem);
 352   }
 353   transform_later(*mem);
 354 }
 355 
 356 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
 357                                               Node* src_offset,
 358                                               Node* dest_offset,
 359                                               bool disjoint_bases,
 360                                               const char* &name,
 361                                               bool dest_uninitialized) {
 362   const TypeInt* src_offset_inttype  = _igvn.find_int_type(src_offset);
 363   const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);
 364 
 365   bool aligned = false;
 366   bool disjoint = disjoint_bases;
 367 
 368   // if the offsets are the same, we can treat the memory regions as
 369   // disjoint, because either the memory regions are in different arrays,
 370   // or they are identical (which we can treat as disjoint.)  We can also
 371   // treat a copy with a destination index  less that the source index
 372   // as disjoint since a low->high copy will work correctly in this case.
 373   if (src_offset_inttype != nullptr && src_offset_inttype->is_con() &&
 374       dest_offset_inttype != nullptr && dest_offset_inttype->is_con()) {
 375     // both indices are constants
 376     int s_offs = src_offset_inttype->get_con();
 377     int d_offs = dest_offset_inttype->get_con();
 378     int element_size = type2aelembytes(t);
 379     aligned = ((arrayOopDesc::base_offset_in_bytes(t) + (uint)s_offs * element_size) % HeapWordSize == 0) &&
 380               ((arrayOopDesc::base_offset_in_bytes(t) + (uint)d_offs * element_size) % HeapWordSize == 0);
 381     if (s_offs >= d_offs)  disjoint = true;
 382   } else if (src_offset == dest_offset && src_offset != nullptr) {
 383     // This can occur if the offsets are identical non-constants.
 384     disjoint = true;
 385   }
 386 
 387   return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
 388 }
 389 
 390 #define XTOP LP64_ONLY(COMMA top())
 391 
 392 // Generate an optimized call to arraycopy.
 393 // Caller must guard against non-arrays.
 394 // Caller must determine a common array basic-type for both arrays.
 395 // Caller must validate offsets against array bounds.
 396 // The slow_region has already collected guard failure paths
 397 // (such as out of bounds length or non-conformable array types).
 398 // The generated code has this shape, in general:
 399 //
 400 //     if (length == 0)  return   // via zero_path
 401 //     slowval = -1
 402 //     if (types unknown) {
 403 //       slowval = call generic copy loop
 404 //       if (slowval == 0)  return  // via checked_path
 405 //     } else if (indexes in bounds) {
 406 //       if ((is object array) && !(array type check)) {
 407 //         slowval = call checked copy loop
 408 //         if (slowval == 0)  return  // via checked_path
 409 //       } else {
 410 //         call bulk copy loop
 411 //         return  // via fast_path
 412 //       }
 413 //     }
 414 //     // adjust params for remaining work:
 415 //     if (slowval != -1) {
 416 //       n = -1^slowval; src_offset += n; dest_offset += n; length -= n
 417 //     }
 418 //   slow_region:
 419 //     call slow arraycopy(src, src_offset, dest, dest_offset, length)
 420 //     return  // via slow_call_path
 421 //
 422 // This routine is used from several intrinsics:  System.arraycopy,
 423 // Object.clone (the array subcase), and Arrays.copyOf[Range].
 424 //
 425 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
 426                                            Node** ctrl, MergeMemNode* mem, Node** io,
 427                                            const TypePtr* adr_type,
 428                                            BasicType basic_elem_type,
 429                                            Node* src,  Node* src_offset,
 430                                            Node* dest, Node* dest_offset,
 431                                            Node* copy_length,
 432                                            Node* dest_length,
 433                                            bool raw_base,
 434                                            bool disjoint_bases,
 435                                            bool length_never_negative,
 436                                            RegionNode* slow_region) {
 437   Node* orig_dest = dest;
 438   if (slow_region == nullptr) {
 439     slow_region = new RegionNode(1);
 440     transform_later(slow_region);
 441   }
 442 
 443   bool  dest_needs_zeroing   = false;
 444   bool  acopy_to_uninitialized = false;
 445   Node* init_value = nullptr;
 446   Node* raw_init_value = nullptr;
 447 
 448   // See if this is the initialization of a newly-allocated array.
 449   // If so, we will take responsibility here for initializing it to zero.
 450   // (Note:  Because tightly_coupled_allocation performs checks on the
 451   // out-edges of the dest, we need to avoid making derived pointers
 452   // from it until we have checked its uses.)
 453   if (ReduceBulkZeroing
 454       && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
 455       && basic_elem_type != T_CONFLICT // avoid corner case
 456       && !src->eqv_uncast(dest)
 457       && alloc != nullptr
 458       && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
 459     assert(ac->is_alloc_tightly_coupled(), "sanity");
 460     // acopy to uninitialized tightly coupled allocations
 461     // needs zeroing outside the copy range
 462     // and the acopy itself will be to uninitialized memory
 463     acopy_to_uninitialized = true;
 464     if (alloc->maybe_set_complete(&_igvn)) {
 465       // "You break it, you buy it."
 466       InitializeNode* init = alloc->initialization();
 467       assert(init->is_complete(), "we just did this");
 468       init->set_complete_with_arraycopy();
 469       assert(dest->is_CheckCastPP(), "sanity");
 470       assert(dest->in(0)->in(0) == init, "dest pinned");
 471       adr_type = TypeRawPtr::BOTTOM;  // all initializations are into raw memory
 472       raw_base = true;
 473       dest = dest->in(1); // writing to raw memory requires a raw base
 474       // From this point on, every exit path is responsible for
 475       // initializing any non-copied parts of the object to zero.
 476       // Also, if this flag is set we make sure that arraycopy interacts properly
 477       // with G1, eliding pre-barriers. See CR 6627983.
 478       dest_needs_zeroing = true;
 479       init_value = alloc->in(AllocateNode::InitValue);
 480       raw_init_value = alloc->in(AllocateNode::RawInitValue);
 481     } else {
 482       // dest_need_zeroing = false;
 483     }
 484   } else {
 485     // No zeroing elimination needed here.
 486     alloc                  = nullptr;
 487     acopy_to_uninitialized = false;
 488     //dest_needs_zeroing   = false;
 489   }
 490 
 491   uint alias_idx = C->get_alias_index(adr_type);
 492 
 493   // Results are placed here:
 494   enum { fast_path        = 1,  // normal void-returning assembly stub
 495          checked_path     = 2,  // special assembly stub with cleanup
 496          slow_call_path   = 3,  // something went wrong; call the VM
 497          zero_path        = 4,  // bypass when length of copy is zero
 498          bcopy_path       = 5,  // copy primitive array by 64-bit blocks
 499          PATH_LIMIT       = 6
 500   };
 501   RegionNode* result_region = new RegionNode(PATH_LIMIT);
 502   PhiNode*    result_i_o    = new PhiNode(result_region, Type::ABIO);
 503   PhiNode*    result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
 504   assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
 505   transform_later(result_region);
 506   transform_later(result_i_o);
 507   transform_later(result_memory);
 508 
 509   // The slow_control path:
 510   Node* slow_control;
 511   Node* slow_i_o = *io;
 512   Node* slow_mem = mem->memory_at(alias_idx);
 513   DEBUG_ONLY(slow_control = (Node*) badAddress);
 514 
 515   // Checked control path:
 516   Node* checked_control = top();
 517   Node* checked_mem     = nullptr;
 518   Node* checked_i_o     = nullptr;
 519   Node* checked_value   = nullptr;
 520 
 521   if (basic_elem_type == T_CONFLICT) {
 522     assert(!dest_needs_zeroing, "");
 523     Node* cv = generate_generic_arraycopy(ctrl, &mem,
 524                                           adr_type,
 525                                           src, src_offset, dest, dest_offset,
 526                                           copy_length, acopy_to_uninitialized);
 527     if (cv == nullptr)  cv = intcon(-1);  // failure (no stub available)
 528     checked_control = *ctrl;
 529     checked_i_o     = *io;
 530     checked_mem     = mem->memory_at(alias_idx);
 531     checked_value   = cv;
 532     *ctrl = top();
 533   }
 534 
 535   Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
 536   if (not_pos != nullptr) {
 537     Node* local_ctrl = not_pos, *local_io = *io;
 538     MergeMemNode* local_mem = MergeMemNode::make(mem);
 539     transform_later(local_mem);
 540 
 541     // (6) length must not be negative.
 542     if (!length_never_negative) {
 543       generate_negative_guard(&local_ctrl, copy_length, slow_region);
 544     }
 545 
 546     // copy_length is 0.
 547     if (dest_needs_zeroing) {
 548       assert(!local_ctrl->is_top(), "no ctrl?");
 549       if (copy_length->eqv_uncast(dest_length)
 550           || _igvn.find_int_con(dest_length, 1) <= 0) {
 551         // There is no zeroing to do. No need for a secondary raw memory barrier.
 552       } else {
 553         // Clear the whole thing since there are no source elements to copy.
 554         generate_clear_array(local_ctrl, local_mem,
 555                              adr_type, dest,
 556                              init_value, raw_init_value,
 557                              basic_elem_type,
 558                              intcon(0), nullptr,
 559                              alloc->in(AllocateNode::AllocSize), raw_base);
 560         // Use a secondary InitializeNode as raw memory barrier.
 561         // Currently it is needed only on this path since other
 562         // paths have stub or runtime calls as raw memory barriers.
 563         MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
 564                                           Compile::AliasIdxRaw,
 565                                           top());
 566         transform_later(mb);
 567         mb->set_req(TypeFunc::Control,local_ctrl);
 568         mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
 569         local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
 570         local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
 571 
 572         InitializeNode* init = mb->as_Initialize();
 573         init->set_complete(&_igvn);  // (there is no corresponding AllocateNode)
 574       }
 575     }
 576 
 577     // Present the results of the fast call.
 578     result_region->init_req(zero_path, local_ctrl);
 579     result_i_o   ->init_req(zero_path, local_io);
 580     result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
 581   }
 582 
 583   if (!(*ctrl)->is_top() && dest_needs_zeroing) {
 584     // We have to initialize the *uncopied* part of the array to zero.
 585     // The copy destination is the slice dest[off..off+len].  The other slices
 586     // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
 587     Node* dest_size   = alloc->in(AllocateNode::AllocSize);
 588     Node* dest_tail   = transform_later( new AddINode(dest_offset, copy_length));
 589 
 590     // If there is a head section that needs zeroing, do it now.
 591     if (_igvn.find_int_con(dest_offset, -1) != 0) {
 592       generate_clear_array(*ctrl, mem,
 593                            adr_type, dest,
 594                            init_value, raw_init_value,
 595                            basic_elem_type,
 596                            intcon(0), dest_offset,
 597                            nullptr, raw_base);
 598     }
 599 
 600     // Next, perform a dynamic check on the tail length.
 601     // It is often zero, and we can win big if we prove this.
 602     // There are two wins:  Avoid generating the ClearArray
 603     // with its attendant messy index arithmetic, and upgrade
 604     // the copy to a more hardware-friendly word size of 64 bits.
 605     Node* tail_ctl = nullptr;
 606     if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
 607       Node* cmp_lt   = transform_later( new CmpINode(dest_tail, dest_length) );
 608       Node* bol_lt   = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
 609       tail_ctl = generate_slow_guard(ctrl, bol_lt, nullptr);
 610       assert(tail_ctl != nullptr || !(*ctrl)->is_top(), "must be an outcome");
 611     }
 612 
 613     // At this point, let's assume there is no tail.
 614     if (!(*ctrl)->is_top() && alloc != nullptr && basic_elem_type != T_OBJECT) {
 615       // There is no tail.  Try an upgrade to a 64-bit copy.
 616       bool didit = false;
 617       {
 618         Node* local_ctrl = *ctrl, *local_io = *io;
 619         MergeMemNode* local_mem = MergeMemNode::make(mem);
 620         transform_later(local_mem);
 621 
 622         didit = generate_block_arraycopy(&local_ctrl, &local_mem, adr_type,
 623                                          basic_elem_type, src, src_offset,
 624                                          dest, dest_offset, dest_size, acopy_to_uninitialized, raw_base);
 625         if (didit) {
 626           // Present the results of the block-copying fast call.
 627           result_region->init_req(bcopy_path, local_ctrl);
 628           result_i_o   ->init_req(bcopy_path, local_io);
 629           result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
 630         }
 631       }
 632       if (didit) {
 633         *ctrl = top();     // no regular fast path
 634       }
 635     }
 636 
 637     // Clear the tail, if any.
 638     if (tail_ctl != nullptr) {
 639       Node* notail_ctl = (*ctrl)->is_top() ? nullptr : *ctrl;
 640       *ctrl = tail_ctl;
 641       if (notail_ctl == nullptr) {
 642         generate_clear_array(*ctrl, mem,
 643                              adr_type, dest,
 644                              init_value, raw_init_value,
 645                              basic_elem_type,
 646                              dest_tail, nullptr,
 647                              dest_size, raw_base);
 648       } else {
 649         // Make a local merge.
 650         Node* done_ctl = transform_later(new RegionNode(3));
 651         Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
 652         done_ctl->init_req(1, notail_ctl);
 653         done_mem->init_req(1, mem->memory_at(alias_idx));
 654         generate_clear_array(*ctrl, mem,
 655                              adr_type, dest,
 656                              init_value, raw_init_value,
 657                              basic_elem_type,
 658                              dest_tail, nullptr,
 659                              dest_size, raw_base);
 660         done_ctl->init_req(2, *ctrl);
 661         done_mem->init_req(2, mem->memory_at(alias_idx));
 662         *ctrl = done_ctl;
 663         mem->set_memory_at(alias_idx, done_mem);
 664       }
 665     }
 666   }
 667 
 668   BasicType copy_type = basic_elem_type;
 669   assert(basic_elem_type != T_ARRAY, "caller must fix this");
 670   if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
 671     // If src and dest have compatible element types, we can copy bits.
 672     // Types S[] and D[] are compatible if D is a supertype of S.
 673     //
 674     // If they are not, we will use checked_oop_disjoint_arraycopy,
 675     // which performs a fast optimistic per-oop check, and backs off
 676     // further to JVM_ArrayCopy on the first per-oop check that fails.
 677     // (Actually, we don't move raw bits only; the GC requires card marks.)
 678 
 679     // We don't need a subtype check for validated copies and Object[].clone()
 680     bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
 681                               ac->is_copyofrange_validated() || ac->is_clone_oop_array();
 682     if (!skip_subtype_check) {
 683       // Get the klass* for both src and dest
 684       Node* src_klass  = ac->in(ArrayCopyNode::SrcKlass);
 685       Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
 686 
 687       assert(src_klass != nullptr && dest_klass != nullptr, "should have klasses");
 688 
 689       // Generate the subtype check.
 690       // This might fold up statically, or then again it might not.
 691       //
 692       // Non-static example:  Copying List<String>.elements to a new String[].
 693       // The backing store for a List<String> is always an Object[],
 694       // but its elements are always type String, if the generic types
 695       // are correct at the source level.
 696       //
 697       // Test S[] against D[], not S against D, because (probably)
 698       // the secondary supertype cache is less busy for S[] than S.
 699       // This usually only matters when D is an interface.
 700       Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, _igvn, nullptr, -1);
 701       // Plug failing path into checked_oop_disjoint_arraycopy
 702       if (not_subtype_ctrl != top()) {
 703         Node* local_ctrl = not_subtype_ctrl;
 704         MergeMemNode* local_mem = MergeMemNode::make(mem);
 705         transform_later(local_mem);
 706 
 707         // (At this point we can assume disjoint_bases, since types differ.)
 708         int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
 709         Node* p1 = off_heap_plus_addr(dest_klass, ek_offset);
 710         Node* n1 = LoadKlassNode::make(_igvn, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
 711         Node* dest_elem_klass = transform_later(n1);
 712         Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
 713                                                 adr_type,
 714                                                 dest_elem_klass,
 715                                                 src, src_offset, dest, dest_offset,
 716                                                 ConvI2X(copy_length), acopy_to_uninitialized, raw_base);
 717         if (cv == nullptr)  cv = intcon(-1);  // failure (no stub available)
 718         checked_control = local_ctrl;
 719         checked_i_o     = *io;
 720         checked_mem     = local_mem->memory_at(alias_idx);
 721         checked_value   = cv;
 722       }
 723     }
 724     // At this point we know we do not need type checks on oop stores.
 725 
 726     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 727     if (!bs->array_copy_requires_gc_barriers(alloc != nullptr, copy_type, false, false, BarrierSetC2::Expansion)) {
 728       // If we do not need gc barriers, copy using the jint or jlong stub.
 729       copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
 730       assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
 731              "sizes agree");
 732     }
 733   }
 734 
 735   if (!(*ctrl)->is_top()) {
 736     // Generate the fast path, if possible.
 737     Node* local_ctrl = *ctrl;
 738     MergeMemNode* local_mem = MergeMemNode::make(mem);
 739     transform_later(local_mem);
 740     generate_unchecked_arraycopy(&local_ctrl, &local_mem,
 741                                  adr_type, copy_type, disjoint_bases,
 742                                  src, src_offset, dest, dest_offset,
 743                                  ConvI2X(copy_length), acopy_to_uninitialized,
 744                                  raw_base);
 745 
 746     // Present the results of the fast call.
 747     result_region->init_req(fast_path, local_ctrl);
 748     result_i_o   ->init_req(fast_path, *io);
 749     result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
 750   }
 751 
 752   // Here are all the slow paths up to this point, in one bundle:
 753   assert(slow_region != nullptr, "allocated on entry");
 754   slow_control = slow_region;
 755   DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
 756 
 757   *ctrl = checked_control;
 758   if (!(*ctrl)->is_top()) {
 759     // Clean up after the checked call.
 760     // The returned value is either 0 or -1^K,
 761     // where K = number of partially transferred array elements.
 762     Node* cmp = new CmpINode(checked_value, intcon(0));
 763     transform_later(cmp);
 764     Node* bol = new BoolNode(cmp, BoolTest::eq);
 765     transform_later(bol);
 766     IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
 767     transform_later(iff);
 768 
 769     // If it is 0, we are done, so transfer to the end.
 770     Node* checks_done = new IfTrueNode(iff);
 771     transform_later(checks_done);
 772     result_region->init_req(checked_path, checks_done);
 773     result_i_o   ->init_req(checked_path, checked_i_o);
 774     result_memory->init_req(checked_path, checked_mem);
 775 
 776     // If it is not zero, merge into the slow call.
 777     *ctrl = new IfFalseNode(iff);
 778     transform_later(*ctrl);
 779     RegionNode* slow_reg2 = new RegionNode(3);
 780     PhiNode*    slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
 781     PhiNode*    slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
 782     transform_later(slow_reg2);
 783     transform_later(slow_i_o2);
 784     transform_later(slow_mem2);
 785     slow_reg2  ->init_req(1, slow_control);
 786     slow_i_o2  ->init_req(1, slow_i_o);
 787     slow_mem2  ->init_req(1, slow_mem);
 788     slow_reg2  ->init_req(2, *ctrl);
 789     slow_i_o2  ->init_req(2, checked_i_o);
 790     slow_mem2  ->init_req(2, checked_mem);
 791 
 792     slow_control = slow_reg2;
 793     slow_i_o     = slow_i_o2;
 794     slow_mem     = slow_mem2;
 795 
 796     if (alloc != nullptr) {
 797       // We'll restart from the very beginning, after zeroing the whole thing.
 798       // This can cause double writes, but that's OK since dest is brand new.
 799       // So we ignore the low 31 bits of the value returned from the stub.
 800     } else {
 801       // We must continue the copy exactly where it failed, or else
 802       // another thread might see the wrong number of writes to dest.
 803       Node* checked_offset = new XorINode(checked_value, intcon(-1));
 804       Node* slow_offset    = new PhiNode(slow_reg2, TypeInt::INT);
 805       transform_later(checked_offset);
 806       transform_later(slow_offset);
 807       slow_offset->init_req(1, intcon(0));
 808       slow_offset->init_req(2, checked_offset);
 809 
 810       // Adjust the arguments by the conditionally incoming offset.
 811       Node* src_off_plus  = new AddINode(src_offset,  slow_offset);
 812       transform_later(src_off_plus);
 813       Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
 814       transform_later(dest_off_plus);
 815       Node* length_minus  = new SubINode(copy_length, slow_offset);
 816       transform_later(length_minus);
 817 
 818       // Tweak the node variables to adjust the code produced below:
 819       src_offset  = src_off_plus;
 820       dest_offset = dest_off_plus;
 821       copy_length = length_minus;
 822     }
 823   }
 824   *ctrl = slow_control;
 825   if (!(*ctrl)->is_top()) {
 826     Node* local_ctrl = *ctrl, *local_io = slow_i_o;
 827     MergeMemNode* local_mem = MergeMemNode::make(mem);
 828     transform_later(local_mem);
 829 
 830     // Generate the slow path, if needed.
 831     local_mem->set_memory_at(alias_idx, slow_mem);
 832 
 833     if (dest_needs_zeroing) {
 834       generate_clear_array(local_ctrl, local_mem,
 835                            adr_type, dest,
 836                            init_value, raw_init_value,
 837                            basic_elem_type,
 838                            intcon(0), nullptr,
 839                            alloc->in(AllocateNode::AllocSize),
 840                            raw_base);
 841     }
 842 
 843     local_mem = generate_slow_arraycopy(ac,
 844                                         &local_ctrl, local_mem, &local_io,
 845                                         adr_type,
 846                                         src, src_offset, orig_dest, dest_offset,
 847                                         copy_length, /*dest_uninitialized*/false);
 848 
 849     result_region->init_req(slow_call_path, local_ctrl);
 850     result_i_o   ->init_req(slow_call_path, local_io);
 851     result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
 852   } else {
 853     ShouldNotReachHere(); // no call to generate_slow_arraycopy:
 854                           // projections were not extracted
 855   }
 856 
 857   // Remove unused edges.
 858   for (uint i = 1; i < result_region->req(); i++) {
 859     if (result_region->in(i) == nullptr) {
 860       result_region->init_req(i, top());
 861     }
 862   }
 863 
 864   // Finished; return the combined state.
 865   *ctrl = result_region;
 866   *io = result_i_o;
 867   mem->set_memory_at(alias_idx, result_memory);
 868 
 869   // mem no longer guaranteed to stay a MergeMemNode
 870   Node* out_mem = mem;
 871   DEBUG_ONLY(mem = nullptr);
 872 
 873   // The memory edges above are precise in order to model effects around
 874   // array copies accurately to allow value numbering of field loads around
 875   // arraycopy.  Such field loads, both before and after, are common in Java
 876   // collections and similar classes involving header/array data structures.
 877   //
 878   // But with low number of register or when some registers are used or killed
 879   // by arraycopy calls it causes registers spilling on stack. See 6544710.
 880   // The next memory barrier is added to avoid it. If the arraycopy can be
 881   // optimized away (which it can, sometimes) then we can manually remove  // Do not let reads from the cloned object float above the arrayco
 882   // the membar also.
 883   //
 884   // Do not let reads from the cloned object float above the arraycopy.
 885   if (alloc != nullptr && !alloc->initialization()->does_not_escape()) {
 886     // Do not let stores that initialize this object be reordered with
 887     // a subsequent store that would make this object accessible by
 888     // other threads.
 889     assert(ac->_dest_type == TypeOopPtr::BOTTOM, "non escaping destination shouldn't have narrow slice");
 890     insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore, Compile::AliasIdxBot);
 891   } else {
 892     int mem_bar_alias_idx = Compile::AliasIdxBot;
 893     if (ac->_dest_type != TypeOopPtr::BOTTOM) {
 894       // The graph was transformed under the assumption the ArrayCopy node only had an effect on a narrow slice. We can't
 895       // insert a wide membar now that it's being expanded: a load that uses the input memory state of the ArrayCopy
 896       // could then become anti dependent on the membar when it was not anti dependent on the ArrayCopy leading to a
 897       // broken graph.
 898       mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
 899     }
 900     insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
 901   }
 902 
 903   assert((*ctrl)->is_Proj(), "MemBar control projection");
 904   assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
 905   (*ctrl)->in(0)->isa_MemBar()->set_trailing_expanded_array_copy();
 906 
 907   _igvn.replace_node(_callprojs->fallthrough_memproj, out_mem);
 908   if (_callprojs->fallthrough_ioproj != nullptr) {
 909     _igvn.replace_node(_callprojs->fallthrough_ioproj, *io);
 910   }
 911   _igvn.replace_node(_callprojs->fallthrough_catchproj, *ctrl);
 912 
 913 #ifdef ASSERT
 914   const TypeOopPtr* dest_t = _igvn.type(orig_dest)->is_oopptr();
 915   if (dest_t->is_known_instance()) {
 916     ArrayCopyNode* ac = nullptr;
 917     assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
 918     assert(ac == nullptr, "no arraycopy anymore");
 919   }
 920 #endif
 921 
 922   return out_mem;
 923 }
 924 
 925 // Helper for initialization of arrays, creating a ClearArray.
 926 // It writes zero bits in [start..end), within the body of an array object.
 927 // The memory effects are all chained onto the 'adr_type' alias category.
 928 //
 929 // Since the object is otherwise uninitialized, we are free
 930 // to put a little "slop" around the edges of the cleared area,
 931 // as long as it does not go back into the array's header,
 932 // or beyond the array end within the heap.
 933 //
 934 // The lower edge can be rounded down to the nearest jint and the
 935 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
 936 //
 937 // Arguments:
 938 //   adr_type           memory slice where writes are generated
 939 //   dest               oop of the destination array
 940 //   basic_elem_type    element type of the destination
 941 //   slice_idx          array index of first element to store
 942 //   slice_len          number of elements to store (or null)
 943 //   dest_size          total size in bytes of the array object
 944 //
 945 // Exactly one of slice_len or dest_size must be non-null.
 946 // If dest_size is non-null, zeroing extends to the end of the object.
 947 // If slice_len is non-null, the slice_idx value must be a constant.
 948 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
 949                                             const TypePtr* adr_type,
 950                                             Node* dest,
 951                                             Node* val,
 952                                             Node* raw_val,
 953                                             BasicType basic_elem_type,
 954                                             Node* slice_idx,
 955                                             Node* slice_len,
 956                                             Node* dest_size,
 957                                             const bool raw_base) {
 958   // one or the other but not both of slice_len and dest_size:
 959   assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
 960   if (slice_len == nullptr)  slice_len = top();
 961   if (dest_size == nullptr)  dest_size = top();
 962 
 963   uint alias_idx = C->get_alias_index(adr_type);
 964 
 965   // operate on this memory slice:
 966   Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
 967 
 968   // scaling and rounding of indexes:
 969   int scale = exact_log2(type2aelembytes(basic_elem_type));
 970   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 971   int clear_low = (-1 << scale) & (BytesPerInt  - 1);
 972   int bump_bit  = (-1 << scale) & BytesPerInt;
 973 
 974   // determine constant starts and ends
 975   const intptr_t BIG_NEG = -128;
 976   assert(BIG_NEG + 2*abase < 0, "neg enough");
 977   intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
 978   intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
 979   if (slice_len_con == 0) {
 980     return;                     // nothing to do here
 981   }
 982   intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
 983   intptr_t end_con   = _igvn.find_intptr_t_con(dest_size, -1);
 984   if (slice_idx_con >= 0 && slice_len_con >= 0) {
 985     assert(end_con < 0, "not two cons");
 986     end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
 987                        BytesPerLong);
 988   }
 989 
 990   if (start_con >= 0 && end_con >= 0) {
 991     // Constant start and end.  Simple.
 992     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
 993                                        start_con, end_con, raw_base, &_igvn);
 994   } else if (start_con >= 0 && dest_size != top()) {
 995     // Constant start, pre-rounded end after the tail of the array.
 996     Node* end = dest_size;
 997     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
 998                                        start_con, end, raw_base, &_igvn);
 999   } else if (start_con >= 0 && slice_len != top()) {
1000     // Constant start, non-constant end.  End needs rounding up.
1001     // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
1002     intptr_t end_base  = abase + (slice_idx_con << scale);
1003     int      end_round = (-1 << scale) & (BytesPerLong  - 1);
1004     Node*    end       = ConvI2X(slice_len);
1005     if (scale != 0)
1006       end = transform_later(new LShiftXNode(end, intcon(scale) ));
1007     end_base += end_round;
1008     end = transform_later(new AddXNode(end, MakeConX(end_base)) );
1009     end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
1010     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, val, raw_val,
1011                                        start_con, end, raw_base, &_igvn);
1012   } else if (start_con < 0 && dest_size != top()) {
1013     // Non-constant start, pre-rounded end after the tail of the array.
1014     // This is almost certainly a "round-to-end" operation.
1015     Node* start = slice_idx;
1016     start = ConvI2X(start);
1017     if (scale != 0)
1018       start = transform_later(new LShiftXNode( start, intcon(scale) ));
1019     start = transform_later(new AddXNode(start, MakeConX(abase)) );
1020     if ((bump_bit | clear_low) != 0) {
1021       int to_clear = (bump_bit | clear_low);
1022       // Align up mod 8, then store a jint zero unconditionally
1023       // just before the mod-8 boundary.
1024       if (((abase + bump_bit) & ~to_clear) - bump_bit
1025           < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
1026         bump_bit = 0;
1027         assert((abase & to_clear) == 0, "array base must be long-aligned");
1028       } else {
1029         // Bump 'start' up to (or past) the next jint boundary:
1030         start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
1031         assert((abase & clear_low) == 0, "array base must be int-aligned");
1032       }
1033       // Round bumped 'start' down to jlong boundary in body of array.
1034       start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
1035       if (bump_bit != 0) {
1036         // Store a zero to the immediately preceding jint:
1037         Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
1038         Node* p1 = basic_plus_adr(dest, x1, raw_base);
1039         if (val == nullptr) {
1040           assert(raw_val == nullptr, "val may not be null");
1041           mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
1042         } else {
1043           assert(_igvn.type(val)->isa_narrowoop(), "should be narrow oop");
1044           mem = new StoreNNode(ctrl, mem, p1, adr_type, val, MemNode::unordered);
1045         }
1046         mem = transform_later(mem);
1047       }
1048     }
1049     Node* end = dest_size; // pre-rounded
1050     mem = ClearArrayNode::clear_memory(ctrl, mem, dest, raw_val,
1051                                        start, end, raw_base, &_igvn);
1052   } else {
1053     // Non-constant start, unrounded non-constant end.
1054     // (Nobody zeroes a random midsection of an array using this routine.)
1055     ShouldNotReachHere();       // fix caller
1056   }
1057 
1058   // Done.
1059   merge_mem->set_memory_at(alias_idx, mem);
1060 }
1061 
1062 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem,
1063                                                 const TypePtr* adr_type,
1064                                                 BasicType basic_elem_type,
1065                                                 Node* src, Node* src_offset,
1066                                                 Node* dest, Node* dest_offset,
1067                                                 Node* dest_size, bool dest_uninitialized,
1068                                                 const bool raw_base) {
1069   // See if there is an advantage from block transfer.
1070   int scale = exact_log2(type2aelembytes(basic_elem_type));
1071   if (scale >= LogBytesPerLong)
1072     return false;               // it is already a block transfer
1073 
1074   // Look at the alignment of the starting offsets.
1075   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
1076 
1077   intptr_t src_off_con  = (intptr_t) _igvn.find_int_con(src_offset, -1);
1078   intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
1079   if (src_off_con < 0 || dest_off_con < 0) {
1080     // At present, we can only understand constants.
1081     return false;
1082   }
1083 
1084   intptr_t src_off  = abase + (src_off_con  << scale);
1085   intptr_t dest_off = abase + (dest_off_con << scale);
1086 
1087   if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1088     // Non-aligned; too bad.
1089     // One more chance:  Pick off an initial 32-bit word.
1090     // This is a common case, since abase can be odd mod 8.
1091     if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1092         ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1093       Node* sptr = basic_plus_adr(src,  src_off);
1094       Node* dptr = basic_plus_adr(dest, dest_off, raw_base);
1095       const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
1096       assert(s_adr_type->isa_aryptr(), "impossible slice");
1097       uint s_alias_idx = C->get_alias_index(s_adr_type);
1098       uint d_alias_idx = C->get_alias_index(adr_type);
1099       bool is_mismatched = (basic_elem_type != T_INT);
1100       Node* sval = transform_later(
1101           LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
1102                          TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
1103                          false /*require_atomic_access*/, false /*unaligned*/, is_mismatched));
1104       Node* st = transform_later(
1105           StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
1106                           sval, T_INT, MemNode::unordered));
1107       if (is_mismatched) {
1108         st->as_Store()->set_mismatched_access();
1109       }
1110       (*mem)->set_memory_at(d_alias_idx, st);
1111       src_off += BytesPerInt;
1112       dest_off += BytesPerInt;
1113     } else {
1114       return false;
1115     }
1116   }
1117   assert(src_off % BytesPerLong == 0, "");
1118   assert(dest_off % BytesPerLong == 0, "");
1119 
1120   // Do this copy by giant steps.
1121   Node* sptr  = basic_plus_adr(src,  src_off);
1122   Node* dptr  = basic_plus_adr(dest, dest_off, raw_base);
1123   Node* countx = dest_size;
1124   countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
1125   countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
1126 
1127   bool disjoint_bases = true;   // since alloc isn't null
1128   generate_unchecked_arraycopy(ctrl, mem,
1129                                adr_type, T_LONG, disjoint_bases,
1130                                sptr, nullptr, dptr, nullptr, countx, dest_uninitialized, raw_base);
1131 
1132   return true;
1133 }
1134 
1135 // Helper function; generates code for the slow case.
1136 // We make a call to a runtime method which emulates the native method,
1137 // but without the native wrapper overhead.
1138 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
1139                                                         Node** ctrl, Node* mem, Node** io,
1140                                                         const TypePtr* adr_type,
1141                                                         Node* src,  Node* src_offset,
1142                                                         Node* dest, Node* dest_offset,
1143                                                         Node* copy_length, bool dest_uninitialized) {
1144   assert(!dest_uninitialized, "Invariant");
1145 
1146   const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1147   CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1148                                           "slow_arraycopy", TypePtr::BOTTOM);
1149 
1150   call->init_req(TypeFunc::Control, *ctrl);
1151   call->init_req(TypeFunc::I_O    , *io);
1152   call->init_req(TypeFunc::Memory , mem);
1153   call->init_req(TypeFunc::ReturnAdr, top());
1154   call->init_req(TypeFunc::FramePtr, top());
1155   call->init_req(TypeFunc::Parms+0, src);
1156   call->init_req(TypeFunc::Parms+1, src_offset);
1157   call->init_req(TypeFunc::Parms+2, dest);
1158   call->init_req(TypeFunc::Parms+3, dest_offset);
1159   call->init_req(TypeFunc::Parms+4, copy_length);
1160   call->copy_call_debug_info(&_igvn, ac);
1161 
1162   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1163   _igvn.replace_node(ac, call);
1164   transform_later(call);
1165 
1166   _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1167   *ctrl = _callprojs->fallthrough_catchproj->clone();
1168   transform_later(*ctrl);
1169 
1170   Node* m = _callprojs->fallthrough_memproj->clone();
1171   transform_later(m);
1172 
1173   uint alias_idx = C->get_alias_index(adr_type);
1174   MergeMemNode* out_mem;
1175   if (alias_idx != Compile::AliasIdxBot) {
1176     out_mem = MergeMemNode::make(mem);
1177     out_mem->set_memory_at(alias_idx, m);
1178   } else {
1179     out_mem = MergeMemNode::make(m);
1180   }
1181   transform_later(out_mem);
1182 
1183   // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1184   // could be nullptr. Skip clone and update nullptr fallthrough_ioproj.
1185   if (_callprojs->fallthrough_ioproj != nullptr) {
1186     *io = _callprojs->fallthrough_ioproj->clone();
1187     transform_later(*io);
1188   } else {
1189     *io = nullptr;
1190   }
1191 
1192   return out_mem;
1193 }
1194 
1195 // Helper function; generates code for cases requiring runtime checks.
1196 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1197                                                      const TypePtr* adr_type,
1198                                                      Node* dest_elem_klass,
1199                                                      Node* src,  Node* src_offset,
1200                                                      Node* dest, Node* dest_offset,
1201                                                      Node* copy_length, bool dest_uninitialized,
1202                                                      const bool raw_base) {
1203   if ((*ctrl)->is_top())  return nullptr;
1204 
1205   address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1206   if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1207     return nullptr;
1208   }
1209 
1210   // Pick out the parameters required to perform a store-check
1211   // for the target array.  This is an optimistic check.  It will
1212   // look in each non-null element's class, at the desired klass's
1213   // super_check_offset, for the desired klass.
1214   int sco_offset = in_bytes(Klass::super_check_offset_offset());
1215   Node* p3 = off_heap_plus_addr(dest_elem_klass, sco_offset);
1216   Node* n3 = new LoadINode(nullptr, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1217   Node* check_offset = ConvI2X(transform_later(n3));
1218   Node* check_value  = dest_elem_klass;
1219 
1220   Node* src_start  = array_element_address(src,  src_offset,  T_OBJECT, false);
1221   Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT, raw_base);
1222 
1223   const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1224   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1225                               src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1226 
1227   finish_arraycopy_call(call, ctrl, mem, adr_type);
1228 
1229   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1230   transform_later(proj);
1231 
1232   return proj;
1233 }
1234 
1235 // Helper function; generates code for cases requiring runtime checks.
1236 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1237                                                    const TypePtr* adr_type,
1238                                                    Node* src,  Node* src_offset,
1239                                                    Node* dest, Node* dest_offset,
1240                                                    Node* copy_length, bool dest_uninitialized) {
1241   if ((*ctrl)->is_top()) return nullptr;
1242   assert(!dest_uninitialized, "Invariant");
1243 
1244   address copyfunc_addr = StubRoutines::generic_arraycopy();
1245   if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1246     return nullptr;
1247   }
1248 
1249   const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1250   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1251                               src, src_offset, dest, dest_offset, copy_length);
1252 
1253   finish_arraycopy_call(call, ctrl, mem, adr_type);
1254 
1255   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1256   transform_later(proj);
1257 
1258   return proj;
1259 }
1260 
1261 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1262 void PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1263                                                     const TypePtr* adr_type,
1264                                                     BasicType basic_elem_type,
1265                                                     bool disjoint_bases,
1266                                                     Node* src,  Node* src_offset,
1267                                                     Node* dest, Node* dest_offset,
1268                                                     Node* copy_length, bool dest_uninitialized,
1269                                                     const bool raw_base) {
1270   if ((*ctrl)->is_top()) {
1271     return;
1272   }
1273 
1274   Node* src_start  = src;
1275   Node* dest_start = dest;
1276   if (src_offset != nullptr || dest_offset != nullptr) {
1277     src_start =  array_element_address(src, src_offset, basic_elem_type, false);
1278     dest_start = array_element_address(dest, dest_offset, basic_elem_type, raw_base);
1279   }
1280 
1281   // Figure out which arraycopy runtime method to call.
1282   const char* copyfunc_name = "arraycopy";
1283   address     copyfunc_addr =
1284       basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1285                           disjoint_bases, copyfunc_name, dest_uninitialized);
1286 
1287   Node* result_memory = nullptr;
1288   RegionNode* exit_block = nullptr;
1289   if (ArrayOperationPartialInlineSize > 0 && is_subword_type(basic_elem_type) &&
1290     Matcher::vector_width_in_bytes(basic_elem_type) >= 16) {
1291     generate_partial_inlining_block(ctrl, mem, adr_type, &exit_block, &result_memory,
1292                                     copy_length, src_start, dest_start, basic_elem_type);
1293   }
1294 
1295   const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1296   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1297                               src_start, dest_start, copy_length XTOP);
1298 
1299   finish_arraycopy_call(call, ctrl, mem, adr_type);
1300 
1301   // Connecting remaining edges for exit_block coming from stub_block.
1302   if (exit_block) {
1303     exit_block->init_req(2, *ctrl);
1304 
1305     // Memory edge corresponding to stub_region.
1306     result_memory->init_req(2, *mem);
1307 
1308     uint alias_idx = C->get_alias_index(adr_type);
1309     if (alias_idx != Compile::AliasIdxBot) {
1310       *mem = MergeMemNode::make(*mem);
1311       (*mem)->set_memory_at(alias_idx, result_memory);
1312     } else {
1313       *mem = MergeMemNode::make(result_memory);
1314     }
1315     transform_later(*mem);
1316     *ctrl = exit_block;
1317   }
1318 }
1319 
1320 const TypePtr* PhaseMacroExpand::adjust_for_flat_array(const TypeAryPtr* top_dest, Node*& src_offset,
1321                                                        Node*& dest_offset, Node*& length, BasicType& dest_elem,
1322                                                        Node*& dest_length) {
1323 #ifdef ASSERT
1324   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1325   bool needs_barriers = top_dest->elem()->inline_klass()->contains_oops() &&
1326     bs->array_copy_requires_gc_barriers(dest_length != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization);
1327   assert(!needs_barriers || StressReflectiveCode, "Flat arracopy would require GC barriers");
1328 #endif
1329   int elem_size = top_dest->flat_elem_size();
1330   if (elem_size >= 8) {
1331     if (elem_size > 8) {
1332       // treat as array of long but scale length, src offset and dest offset
1333       assert((elem_size % 8) == 0, "not a power of 2?");
1334       int factor = elem_size / 8;
1335       length = transform_later(new MulINode(length, intcon(factor)));
1336       src_offset = transform_later(new MulINode(src_offset, intcon(factor)));
1337       dest_offset = transform_later(new MulINode(dest_offset, intcon(factor)));
1338       if (dest_length != nullptr) {
1339         dest_length = transform_later(new MulINode(dest_length, intcon(factor)));
1340       }
1341       elem_size = 8;
1342     }
1343     dest_elem = T_LONG;
1344   } else if (elem_size == 4) {
1345     dest_elem = T_INT;
1346   } else if (elem_size == 2) {
1347     dest_elem = T_CHAR;
1348   } else if (elem_size == 1) {
1349     dest_elem = T_BYTE;
1350   } else {
1351     ShouldNotReachHere();
1352   }
1353   return TypeRawPtr::BOTTOM;
1354 }
1355 
1356 #undef XTOP
1357 
1358 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1359   Node* ctrl = ac->in(TypeFunc::Control);
1360   Node* io = ac->in(TypeFunc::I_O);
1361   Node* src = ac->in(ArrayCopyNode::Src);
1362   Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1363   Node* dest = ac->in(ArrayCopyNode::Dest);
1364   Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1365   Node* length = ac->in(ArrayCopyNode::Length);
1366   MergeMemNode* merge_mem = nullptr;
1367 
1368   if (ac->is_clonebasic()) {
1369     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1370     bs->clone_at_expansion(this, ac);
1371     return;
1372   } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1373     const Type* src_type = _igvn.type(src);
1374     const Type* dest_type = _igvn.type(dest);
1375     const TypeAryPtr* top_src = src_type->isa_aryptr();
1376     // Note: The destination could have type Object (i.e. non-array) when directly invoking the protected method
1377     //       Object::clone() with reflection on a declared Object that is an array at runtime. top_dest is then null.
1378     const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1379     BasicType dest_elem = T_OBJECT;
1380     if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1381       dest_elem = top_dest->elem()->array_element_basic_type();
1382     }
1383     if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1384 
1385     if (top_src != nullptr && top_src->is_flat()) {
1386       // If src is flat, dest is guaranteed to be flat as well
1387       top_dest = top_src;
1388     }
1389 
1390     AllocateArrayNode* alloc = nullptr;
1391     Node* dest_length = nullptr;
1392     if (ac->is_alloc_tightly_coupled()) {
1393       alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1394       assert(alloc != nullptr, "expect alloc");
1395       dest_length = alloc->in(AllocateNode::ALength);
1396     }
1397 
1398     Node* mem = ac->in(TypeFunc::Memory);
1399     const TypePtr* adr_type = nullptr;
1400     bool raw_base = false;
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         raw_base = true;
1419       }
1420     }
1421     merge_mem = MergeMemNode::make(mem);
1422     transform_later(merge_mem);
1423 
1424     generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1425                        adr_type, dest_elem,
1426                        src, src_offset, dest, dest_offset, length,
1427                        dest_length,
1428                        raw_base, true, ac->has_negative_length_guard());
1429 
1430     return;
1431   }
1432 
1433   AllocateArrayNode* alloc = nullptr;
1434   if (ac->is_alloc_tightly_coupled()) {
1435     alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1436     assert(alloc != nullptr, "expect alloc");
1437   }
1438 
1439   assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1440 
1441   // Compile time checks.  If any of these checks cannot be verified at compile time,
1442   // we do not make a fast path for this call.  Instead, we let the call remain as it
1443   // is.  The checks we choose to mandate at compile time are:
1444   //
1445   // (1) src and dest are arrays.
1446   const Type* src_type = src->Value(&_igvn);
1447   const Type* dest_type = dest->Value(&_igvn);
1448   const TypeAryPtr* top_src = src_type->isa_aryptr();
1449   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1450 
1451   BasicType src_elem = T_CONFLICT;
1452   BasicType dest_elem = T_CONFLICT;
1453 
1454   if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1455     src_elem = top_src->elem()->array_element_basic_type();
1456   }
1457   if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1458     dest_elem = top_dest->elem()->array_element_basic_type();
1459   }
1460   if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1461   if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1462 
1463   if (ac->is_arraycopy_validated() && dest_elem != T_CONFLICT && src_elem == T_CONFLICT) {
1464     src_elem = dest_elem;
1465   }
1466 
1467   if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1468     // Conservatively insert a memory barrier on all memory slices.
1469     // Do not let writes into the source float below the arraycopy.
1470     {
1471       Node* mem = ac->in(TypeFunc::Memory);
1472       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1473 
1474       merge_mem = MergeMemNode::make(mem);
1475       transform_later(merge_mem);
1476     }
1477 
1478     // Call StubRoutines::generic_arraycopy stub.
1479     generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1480                        TypeRawPtr::BOTTOM, T_CONFLICT,
1481                        src, src_offset, dest, dest_offset, length,
1482                        nullptr,
1483                        // If a  negative length guard was generated for the ArrayCopyNode,
1484                        // the length of the array can never be negative.
1485                        true, false, ac->has_negative_length_guard());
1486     return;
1487   }
1488 
1489   assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1490 
1491   // (2) src and dest arrays must have elements of the same BasicType
1492   // Figure out the size and type of the elements we will be copying.
1493   //
1494   // We have no stub to copy flat inline type arrays with oop
1495   // fields if we need to emit write barriers.
1496   //
1497   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1498   if (src_elem != dest_elem || top_src->is_flat() != top_dest->is_flat() || dest_elem == T_VOID ||
1499       (top_src->is_flat() && top_dest->elem()->inline_klass()->contains_oops() &&
1500        bs->array_copy_requires_gc_barriers(alloc != nullptr, T_OBJECT, false, false, BarrierSetC2::Optimization))) {
1501     // The component types are not the same or are not recognized.  Punt.
1502     // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1503     {
1504       Node* mem = ac->in(TypeFunc::Memory);
1505       merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1506     }
1507 
1508     _igvn.replace_node(_callprojs->fallthrough_memproj, merge_mem);
1509     if (_callprojs->fallthrough_ioproj != nullptr) {
1510       _igvn.replace_node(_callprojs->fallthrough_ioproj, io);
1511     }
1512     _igvn.replace_node(_callprojs->fallthrough_catchproj, ctrl);
1513     return;
1514   }
1515 
1516   //---------------------------------------------------------------------------
1517   // We will make a fast path for this call to arraycopy.
1518 
1519   // We have the following tests left to perform:
1520   //
1521   // (3) src and dest must not be null.
1522   // (4) src_offset must not be negative.
1523   // (5) dest_offset must not be negative.
1524   // (6) length must not be negative.
1525   // (7) src_offset + length must not exceed length of src.
1526   // (8) dest_offset + length must not exceed length of dest.
1527   // (9) each element of an oop array must be assignable
1528 
1529   Node* mem = ac->in(TypeFunc::Memory);
1530   if (top_dest->is_flat()) {
1531     // Copy to a flat array modifies multiple memory slices. Conservatively insert a barrier
1532     // on all slices to prevent writes into the source from floating below the arraycopy.
1533     int mem_bar_alias_idx = Compile::AliasIdxBot;
1534     if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1535       mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
1536     }
1537     insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
1538   }
1539   merge_mem = MergeMemNode::make(mem);
1540   transform_later(merge_mem);
1541 
1542   RegionNode* slow_region = new RegionNode(1);
1543   transform_later(slow_region);
1544 
1545   if (!ac->is_arraycopy_validated()) {
1546     // (3) operands must not be null
1547     // We currently perform our null checks with the null_check routine.
1548     // This means that the null exceptions will be reported in the caller
1549     // rather than (correctly) reported inside of the native arraycopy call.
1550     // This should be corrected, given time.  We do our null check with the
1551     // stack pointer restored.
1552     // null checks done library_call.cpp
1553 
1554     // (4) src_offset must not be negative.
1555     generate_negative_guard(&ctrl, src_offset, slow_region);
1556 
1557     // (5) dest_offset must not be negative.
1558     generate_negative_guard(&ctrl, dest_offset, slow_region);
1559 
1560     // (6) length must not be negative (moved to generate_arraycopy()).
1561     // generate_negative_guard(length, slow_region);
1562 
1563     // (7) src_offset + length must not exceed length of src.
1564     Node* alen = ac->in(ArrayCopyNode::SrcLen);
1565     assert(alen != nullptr, "need src len");
1566     generate_limit_guard(&ctrl,
1567                          src_offset, length,
1568                          alen,
1569                          slow_region);
1570 
1571     // (8) dest_offset + length must not exceed length of dest.
1572     alen = ac->in(ArrayCopyNode::DestLen);
1573     assert(alen != nullptr, "need dest len");
1574     generate_limit_guard(&ctrl,
1575                          dest_offset, length,
1576                          alen,
1577                          slow_region);
1578 
1579     // (9) each element of an oop array must be assignable
1580     // The generate_arraycopy subroutine checks this.
1581 
1582     // TODO 8350865 This is too strong
1583     // 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.
1584     if (!(top_src->is_flat() && top_dest->is_flat() && top_src->is_null_free() == top_dest->is_null_free())) {
1585       generate_flat_array_guard(&ctrl, src, merge_mem, slow_region);
1586       generate_flat_array_guard(&ctrl, dest, merge_mem, slow_region);
1587       generate_null_free_array_guard(&ctrl, dest, merge_mem, slow_region);
1588     }
1589   }
1590 
1591   // This is where the memory effects are placed:
1592   const TypePtr* adr_type = nullptr;
1593   Node* dest_length = (alloc != nullptr) ? alloc->in(AllocateNode::ALength) : nullptr;
1594   if (top_src->is_flat() && top_dest->is_flat() &&
1595       top_src->is_null_free() == top_dest->is_null_free()) {
1596     adr_type = adjust_for_flat_array(top_dest, src_offset, dest_offset, length, dest_elem, dest_length);
1597   } else if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1598     adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1599   } else {
1600     adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1601   }
1602 
1603   generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1604                      adr_type, dest_elem,
1605                      src, src_offset, dest, dest_offset, length,
1606                      dest_length,
1607                      // If a  negative length guard was generated for the ArrayCopyNode,
1608                      // the length of the array can never be negative.
1609                      false, false, ac->has_negative_length_guard(),
1610                      slow_region);
1611 }