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