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