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