1 /*
   2  * Copyright (c) 2012, 2025, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 

  25 #include "gc/shared/barrierSet.hpp"
  26 #include "gc/shared/tlab_globals.hpp"
  27 #include "oops/objArrayKlass.hpp"
  28 #include "opto/arraycopynode.hpp"
  29 #include "opto/castnode.hpp"
  30 #include "opto/convertnode.hpp"
  31 #include "opto/graphKit.hpp"
  32 #include "opto/macro.hpp"
  33 #include "opto/runtime.hpp"
  34 #include "opto/vectornode.hpp"
  35 #include "runtime/stubRoutines.hpp"
  36 #include "utilities/align.hpp"
  37 #include "utilities/powerOfTwo.hpp"
  38 
  39 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, int alias_idx, Node* precedent) {
  40   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
  41   mb->init_req(TypeFunc::Control, *ctrl);
  42   mb->init_req(TypeFunc::Memory, *mem);
  43   transform_later(mb);
  44   *ctrl = new ProjNode(mb,TypeFunc::Control);
  45   transform_later(*ctrl);
  46   Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
  47   transform_later(mem_proj);
  48   if (alias_idx == Compile::AliasIdxBot) {
  49     *mem = mem_proj;
  50   } else {
  51     MergeMemNode* mm = (*mem)->clone()->as_MergeMem();
  52     mm->set_memory_at(alias_idx, mem_proj);
  53     transform_later(mm);
  54     *mem = mm;
  55   }
  56 }
  57 
  58 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
  59   uint shift  = exact_log2(type2aelembytes(elembt));





  60   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
  61   Node* base =  basic_plus_adr(ary, header);
  62 #ifdef _LP64
  63   // see comment in GraphKit::array_element_address
  64   int index_max = max_jint - 1;  // array size is max_jint, index is one less
  65   const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
  66   idx = transform_later( new ConvI2LNode(idx, lidxtype) );
  67 #endif
  68   Node* scale = new LShiftXNode(idx, intcon(shift));
  69   transform_later(scale);
  70   return basic_plus_adr(ary, base, scale);
  71 }
  72 
  73 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
  74   return transform_later(new ConvI2LNode(offset));
  75 }
  76 
  77 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
  78                                        const TypeFunc* call_type, address call_addr,
  79                                        const char* call_name,
  80                                        const TypePtr* adr_type,
  81                                        Node* parm0, Node* parm1,
  82                                        Node* parm2, Node* parm3,
  83                                        Node* parm4, Node* parm5,
  84                                        Node* parm6, Node* parm7) {
  85   Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
  86   call->init_req(TypeFunc::Control, ctrl);
  87   call->init_req(TypeFunc::I_O    , top());
  88   call->init_req(TypeFunc::Memory , mem);
  89   call->init_req(TypeFunc::ReturnAdr, top());
  90   call->init_req(TypeFunc::FramePtr, top());
  91 
  92   // Hook each parm in order.  Stop looking at the first null.
  93   if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
  94   if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
  95   if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
  96   if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
  97   if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
  98   if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
  99   if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
 100   if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
 101     /* close each nested if ===> */  } } } } } } } }
 102   assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
 103 
 104   return call;
 105 }
 106 
 107 
 108 //------------------------------generate_guard---------------------------
 109 // Helper function for generating guarded fast-slow graph structures.
 110 // The given 'test', if true, guards a slow path.  If the test fails
 111 // then a fast path can be taken.  (We generally hope it fails.)
 112 // In all cases, GraphKit::control() is updated to the fast path.
 113 // The returned value represents the control for the slow path.
 114 // The return value is never 'top'; it is either a valid control
 115 // or null if it is obvious that the slow path can never be taken.
 116 // Also, if region and the slow control are not null, the slow edge
 117 // is appended to the region.
 118 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
 119   if ((*ctrl)->is_top()) {
 120     // Already short circuited.
 121     return nullptr;
 122   }
 123   // Build an if node and its projections.
 124   // If test is true we take the slow path, which we assume is uncommon.
 125   if (_igvn.type(test) == TypeInt::ZERO) {
 126     // The slow branch is never taken.  No need to build this guard.
 127     return nullptr;
 128   }
 129 
 130   IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
 131   transform_later(iff);
 132 
 133   Node* if_slow = new IfTrueNode(iff);
 134   transform_later(if_slow);
 135 
 136   if (region != nullptr) {
 137     region->add_req(if_slow);
 138   }
 139 
 140   Node* if_fast = new IfFalseNode(iff);
 141   transform_later(if_fast);
 142 
 143   *ctrl = if_fast;
 144 
 145   return if_slow;
 146 }
 147 
 148 inline Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
 149   return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
 150 }
 151 




 152 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
 153   if ((*ctrl)->is_top())
 154     return;                // already stopped
 155   if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
 156     return;                // index is already adequately typed
 157   Node* cmp_lt = new CmpINode(index, intcon(0));
 158   transform_later(cmp_lt);
 159   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 160   transform_later(bol_lt);
 161   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 162 }
 163 
 164 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
 165   if ((*ctrl)->is_top())
 166     return;                // already stopped
 167   bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
 168   if (zero_offset && subseq_length->eqv_uncast(array_length))
 169     return;                // common case of whole-array copy
 170   Node* last = subseq_length;
 171   if (!zero_offset) {            // last += offset
 172     last = new AddINode(last, offset);
 173     transform_later(last);
 174   }
 175   Node* cmp_lt = new CmpUNode(array_length, last);
 176   transform_later(cmp_lt);
 177   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 178   transform_later(bol_lt);
 179   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 180 }
 181 
 182 //
 183 // Partial in-lining handling for smaller conjoint/disjoint array copies having
 184 // length(in bytes) less than ArrayOperationPartialInlineSize.
 185 //  if (length <= ArrayOperationPartialInlineSize) {
 186 //    partial_inlining_block:
 187 //      mask = Mask_Gen
 188 //      vload = LoadVectorMasked src , mask
 189 //      StoreVectorMasked dst, mask, vload
 190 //  } else {
 191 //    stub_block:
 192 //      callstub array_copy
 193 //  }
 194 //  exit_block:
 195 //    Phi = label partial_inlining_block:mem , label stub_block:mem (filled by caller)
 196 //    mem = MergeMem (Phi)
 197 //    control = stub_block
 198 //
 199 //  Exit_block and associated phi(memory) are partially initialized for partial_in-lining_block
 200 //  edges. Remaining edges for exit_block coming from stub_block are connected by the caller
 201 //  post stub nodes creation.
 202 //
 203 
 204 void PhaseMacroExpand::generate_partial_inlining_block(Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type,
 205                                                        RegionNode** exit_block, Node** result_memory, Node* length,
 206                                                        Node* src_start, Node* dst_start, BasicType type) {
 207   int inline_limit = ArrayOperationPartialInlineSize / type2aelembytes(type);
 208 
 209   const TypeLong* length_type = _igvn.type(length)->isa_long();
 210   if (length_type == nullptr) {
 211     assert(_igvn.type(length) == Type::TOP, "");
 212     return;
 213   }
 214 
 215   const TypeLong* inline_range = TypeLong::make(0, inline_limit, Type::WidenMin);
 216   if (length_type->join(inline_range) == Type::TOP) {
 217     // The ranges do not intersect, the inline check will surely fail
 218     return;
 219   }
 220 
 221   // Return if the target does not supports masked load/stores.
 222   int lane_count = ArrayCopyNode::get_partial_inline_vector_lane_count(type, length_type->_hi);
 223   if (!Matcher::match_rule_supported_vector(Op_LoadVectorMasked, lane_count, type)  ||
 224       !Matcher::match_rule_supported_vector(Op_StoreVectorMasked, lane_count, type) ||
 225       !Matcher::match_rule_supported_vector(Op_VectorMaskGen, lane_count, type)) {
 226     return;
 227   }
 228 
 229   Node* cmp_le = new CmpULNode(length, longcon(inline_limit));
 230   transform_later(cmp_le);
 231   Node* bol_le = new BoolNode(cmp_le, BoolTest::le);
 232   transform_later(bol_le);
 233   Node* inline_block = generate_guard(ctrl, bol_le, nullptr, PROB_FAIR);
 234   Node* stub_block = *ctrl;
 235 
 236   Node* casted_length = new CastLLNode(inline_block, length, inline_range, ConstraintCastNode::DependencyType::FloatingNarrowing);
 237   transform_later(casted_length);
 238   Node* mask_gen = VectorMaskGenNode::make(casted_length, type);
 239   transform_later(mask_gen);
 240 
 241   unsigned vec_size = lane_count * type2aelembytes(type);
 242   if (C->max_vector_size() < vec_size) {
 243     C->set_max_vector_size(vec_size);
 244   }
 245 
 246   const TypePtr* src_adr_type = _igvn.type(src_start)->isa_ptr();
 247   const TypeVect * vt = TypeVect::make(type, lane_count);
 248   Node* mm = (*mem)->memory_at(C->get_alias_index(src_adr_type));
 249   Node* masked_load = new LoadVectorMaskedNode(inline_block, mm, src_start,
 250                                                src_adr_type, vt, mask_gen);
 251   transform_later(masked_load);
 252 
 253   mm = (*mem)->memory_at(C->get_alias_index(adr_type));
 254   Node* masked_store = new StoreVectorMaskedNode(inline_block, mm, dst_start,
 255                                                  masked_load, adr_type, mask_gen);
 256   transform_later(masked_store);
 257 
 258   // Convergence region for inline_block and stub_block.
 259   *exit_block = new RegionNode(3);
 260   transform_later(*exit_block);
 261   (*exit_block)->init_req(1, inline_block);
 262   *result_memory = new PhiNode(*exit_block, Type::MEMORY, adr_type);
 263   transform_later(*result_memory);
 264   (*result_memory)->init_req(1, masked_store);
 265 
 266   *ctrl = stub_block;
 267 }
 268 
 269 
 270 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
 271   if ((*ctrl)->is_top())  return nullptr;
 272 
 273   if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
 274     return nullptr;                // index is already adequately typed
 275   Node* cmp_le = new CmpINode(index, intcon(0));
 276   transform_later(cmp_le);
 277   BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
 278   Node* bol_le = new BoolNode(cmp_le, le_or_eq);
 279   transform_later(bol_le);
 280   Node* is_notp = generate_guard(ctrl, bol_le, nullptr, PROB_MIN);
 281 
 282   return is_notp;
 283 }
 284 











































 285 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
 286   transform_later(call);
 287 
 288   *ctrl = new ProjNode(call,TypeFunc::Control);
 289   transform_later(*ctrl);
 290   Node* newmem = new ProjNode(call, TypeFunc::Memory);
 291   transform_later(newmem);
 292 
 293   uint alias_idx = C->get_alias_index(adr_type);
 294   if (alias_idx != Compile::AliasIdxBot) {
 295     *mem = MergeMemNode::make(*mem);
 296     (*mem)->set_memory_at(alias_idx, newmem);
 297   } else {
 298     *mem = MergeMemNode::make(newmem);
 299   }
 300   transform_later(*mem);
 301 }
 302 
 303 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
 304                                               Node* src_offset,
 305                                               Node* dest_offset,
 306                                               bool disjoint_bases,
 307                                               const char* &name,
 308                                               bool dest_uninitialized) {
 309   const TypeInt* src_offset_inttype  = _igvn.find_int_type(src_offset);
 310   const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);
 311 
 312   bool aligned = false;
 313   bool disjoint = disjoint_bases;
 314 
 315   // if the offsets are the same, we can treat the memory regions as
 316   // disjoint, because either the memory regions are in different arrays,
 317   // or they are identical (which we can treat as disjoint.)  We can also
 318   // treat a copy with a destination index  less that the source index
 319   // as disjoint since a low->high copy will work correctly in this case.
 320   if (src_offset_inttype != nullptr && src_offset_inttype->is_con() &&
 321       dest_offset_inttype != nullptr && dest_offset_inttype->is_con()) {
 322     // both indices are constants
 323     int s_offs = src_offset_inttype->get_con();
 324     int d_offs = dest_offset_inttype->get_con();
 325     int element_size = type2aelembytes(t);
 326     aligned = ((arrayOopDesc::base_offset_in_bytes(t) + (uint)s_offs * element_size) % HeapWordSize == 0) &&
 327               ((arrayOopDesc::base_offset_in_bytes(t) + (uint)d_offs * element_size) % HeapWordSize == 0);
 328     if (s_offs >= d_offs)  disjoint = true;
 329   } else if (src_offset == dest_offset && src_offset != nullptr) {
 330     // This can occur if the offsets are identical non-constants.
 331     disjoint = true;
 332   }
 333 
 334   return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
 335 }
 336 
 337 #define XTOP LP64_ONLY(COMMA top())
 338 
 339 // Generate an optimized call to arraycopy.
 340 // Caller must guard against non-arrays.
 341 // Caller must determine a common array basic-type for both arrays.
 342 // Caller must validate offsets against array bounds.
 343 // The slow_region has already collected guard failure paths
 344 // (such as out of bounds length or non-conformable array types).
 345 // The generated code has this shape, in general:
 346 //
 347 //     if (length == 0)  return   // via zero_path
 348 //     slowval = -1
 349 //     if (types unknown) {
 350 //       slowval = call generic copy loop
 351 //       if (slowval == 0)  return  // via checked_path
 352 //     } else if (indexes in bounds) {
 353 //       if ((is object array) && !(array type check)) {
 354 //         slowval = call checked copy loop
 355 //         if (slowval == 0)  return  // via checked_path
 356 //       } else {
 357 //         call bulk copy loop
 358 //         return  // via fast_path
 359 //       }
 360 //     }
 361 //     // adjust params for remaining work:
 362 //     if (slowval != -1) {
 363 //       n = -1^slowval; src_offset += n; dest_offset += n; length -= n
 364 //     }
 365 //   slow_region:
 366 //     call slow arraycopy(src, src_offset, dest, dest_offset, length)
 367 //     return  // via slow_call_path
 368 //
 369 // This routine is used from several intrinsics:  System.arraycopy,
 370 // Object.clone (the array subcase), and Arrays.copyOf[Range].
 371 //
 372 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
 373                                            Node** ctrl, MergeMemNode* mem, Node** io,
 374                                            const TypePtr* adr_type,
 375                                            BasicType basic_elem_type,
 376                                            Node* src,  Node* src_offset,
 377                                            Node* dest, Node* dest_offset,
 378                                            Node* copy_length,

 379                                            bool disjoint_bases,
 380                                            bool length_never_negative,
 381                                            RegionNode* slow_region) {
 382   if (slow_region == nullptr) {
 383     slow_region = new RegionNode(1);
 384     transform_later(slow_region);
 385   }
 386 
 387   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, 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   if (!(*ctrl)->is_top()) {
 671     // Generate the fast path, if possible.
 672     Node* local_ctrl = *ctrl;
 673     MergeMemNode* local_mem = MergeMemNode::make(mem);
 674     transform_later(local_mem);
 675     generate_unchecked_arraycopy(&local_ctrl, &local_mem,
 676                                  adr_type, copy_type, disjoint_bases,
 677                                  src, src_offset, dest, dest_offset,
 678                                  ConvI2X(copy_length), acopy_to_uninitialized);
 679 
 680     // Present the results of the fast call.
 681     result_region->init_req(fast_path, local_ctrl);
 682     result_i_o   ->init_req(fast_path, *io);
 683     result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
 684   }
 685 
 686   // Here are all the slow paths up to this point, in one bundle:
 687   assert(slow_region != nullptr, "allocated on entry");
 688   slow_control = slow_region;
 689   DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
 690 
 691   *ctrl = checked_control;
 692   if (!(*ctrl)->is_top()) {
 693     // Clean up after the checked call.
 694     // The returned value is either 0 or -1^K,
 695     // where K = number of partially transferred array elements.
 696     Node* cmp = new CmpINode(checked_value, intcon(0));
 697     transform_later(cmp);
 698     Node* bol = new BoolNode(cmp, BoolTest::eq);
 699     transform_later(bol);
 700     IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
 701     transform_later(iff);
 702 
 703     // If it is 0, we are done, so transfer to the end.
 704     Node* checks_done = new IfTrueNode(iff);
 705     transform_later(checks_done);
 706     result_region->init_req(checked_path, checks_done);
 707     result_i_o   ->init_req(checked_path, checked_i_o);
 708     result_memory->init_req(checked_path, checked_mem);
 709 
 710     // If it is not zero, merge into the slow call.
 711     *ctrl = new IfFalseNode(iff);
 712     transform_later(*ctrl);
 713     RegionNode* slow_reg2 = new RegionNode(3);
 714     PhiNode*    slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
 715     PhiNode*    slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
 716     transform_later(slow_reg2);
 717     transform_later(slow_i_o2);
 718     transform_later(slow_mem2);
 719     slow_reg2  ->init_req(1, slow_control);
 720     slow_i_o2  ->init_req(1, slow_i_o);
 721     slow_mem2  ->init_req(1, slow_mem);
 722     slow_reg2  ->init_req(2, *ctrl);
 723     slow_i_o2  ->init_req(2, checked_i_o);
 724     slow_mem2  ->init_req(2, checked_mem);
 725 
 726     slow_control = slow_reg2;
 727     slow_i_o     = slow_i_o2;
 728     slow_mem     = slow_mem2;
 729 
 730     if (alloc != nullptr) {
 731       // We'll restart from the very beginning, after zeroing the whole thing.
 732       // This can cause double writes, but that's OK since dest is brand new.
 733       // So we ignore the low 31 bits of the value returned from the stub.
 734     } else {
 735       // We must continue the copy exactly where it failed, or else
 736       // another thread might see the wrong number of writes to dest.
 737       Node* checked_offset = new XorINode(checked_value, intcon(-1));
 738       Node* slow_offset    = new PhiNode(slow_reg2, TypeInt::INT);
 739       transform_later(checked_offset);
 740       transform_later(slow_offset);
 741       slow_offset->init_req(1, intcon(0));
 742       slow_offset->init_req(2, checked_offset);
 743 
 744       // Adjust the arguments by the conditionally incoming offset.
 745       Node* src_off_plus  = new AddINode(src_offset,  slow_offset);
 746       transform_later(src_off_plus);
 747       Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
 748       transform_later(dest_off_plus);
 749       Node* length_minus  = new SubINode(copy_length, slow_offset);
 750       transform_later(length_minus);
 751 
 752       // Tweak the node variables to adjust the code produced below:
 753       src_offset  = src_off_plus;
 754       dest_offset = dest_off_plus;
 755       copy_length = length_minus;
 756     }
 757   }
 758   *ctrl = slow_control;
 759   if (!(*ctrl)->is_top()) {
 760     Node* local_ctrl = *ctrl, *local_io = slow_i_o;
 761     MergeMemNode* local_mem = MergeMemNode::make(mem);
 762     transform_later(local_mem);
 763 
 764     // Generate the slow path, if needed.
 765     local_mem->set_memory_at(alias_idx, slow_mem);
 766 
 767     if (dest_needs_zeroing) {
 768       generate_clear_array(local_ctrl, local_mem,
 769                            adr_type, dest, basic_elem_type,


 770                            intcon(0), nullptr,
 771                            alloc->in(AllocateNode::AllocSize));
 772     }
 773 
 774     local_mem = generate_slow_arraycopy(ac,
 775                                         &local_ctrl, local_mem, &local_io,
 776                                         adr_type,
 777                                         src, src_offset, dest, dest_offset,
 778                                         copy_length, /*dest_uninitialized*/false);
 779 
 780     result_region->init_req(slow_call_path, local_ctrl);
 781     result_i_o   ->init_req(slow_call_path, local_io);
 782     result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
 783   } else {
 784     ShouldNotReachHere(); // no call to generate_slow_arraycopy:
 785                           // projections were not extracted
 786   }
 787 
 788   // Remove unused edges.
 789   for (uint i = 1; i < result_region->req(); i++) {
 790     if (result_region->in(i) == nullptr) {
 791       result_region->init_req(i, top());
 792     }
 793   }
 794 
 795   // Finished; return the combined state.
 796   *ctrl = result_region;
 797   *io = result_i_o;
 798   mem->set_memory_at(alias_idx, result_memory);
 799 
 800   // mem no longer guaranteed to stay a MergeMemNode
 801   Node* out_mem = mem;
 802   DEBUG_ONLY(mem = nullptr);
 803 
 804   // The memory edges above are precise in order to model effects around
 805   // array copies accurately to allow value numbering of field loads around
 806   // arraycopy.  Such field loads, both before and after, are common in Java
 807   // collections and similar classes involving header/array data structures.
 808   //
 809   // But with low number of register or when some registers are used or killed
 810   // by arraycopy calls it causes registers spilling on stack. See 6544710.
 811   // The next memory barrier is added to avoid it. If the arraycopy can be
 812   // optimized away (which it can, sometimes) then we can manually remove
 813   // the membar also.
 814   //
 815   // Do not let reads from the cloned object float above the arraycopy.
 816   if (alloc != nullptr && !alloc->initialization()->does_not_escape()) {
 817     // Do not let stores that initialize this object be reordered with
 818     // a subsequent store that would make this object accessible by
 819     // other threads.
 820     assert(ac->_dest_type == TypeOopPtr::BOTTOM, "non escaping destination shouldn't have narrow slice");
 821     insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore, Compile::AliasIdxBot);
 822   } else {
 823     int mem_bar_alias_idx = Compile::AliasIdxBot;
 824     if (ac->_dest_type != TypeOopPtr::BOTTOM) {
 825       // The graph was transformed under the assumption the ArrayCopy node only had an effect on a narrow slice. We can't
 826       // insert a wide membar now that it's being expanded: a load that uses the input memory state of the ArrayCopy
 827       // could then become anti dependent on the membar when it was not anti dependent on the ArrayCopy leading to a
 828       // broken graph.
 829       mem_bar_alias_idx = C->get_alias_index(ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr());
 830     }
 831     insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder, mem_bar_alias_idx);
 832   }
 833 
 834   assert((*ctrl)->is_Proj(), "MemBar control projection");
 835   assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
 836   (*ctrl)->in(0)->isa_MemBar()->set_trailing_expanded_array_copy();
 837 
 838   _igvn.replace_node(_callprojs.fallthrough_memproj, out_mem);
 839   if (_callprojs.fallthrough_ioproj != nullptr) {
 840     _igvn.replace_node(_callprojs.fallthrough_ioproj, *io);
 841   }
 842   _igvn.replace_node(_callprojs.fallthrough_catchproj, *ctrl);
 843 
 844 #ifdef ASSERT
 845   const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
 846   if (dest_t->is_known_instance()) {
 847     ArrayCopyNode* ac = nullptr;
 848     assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
 849     assert(ac == nullptr, "no arraycopy anymore");
 850   }
 851 #endif
 852 
 853   return out_mem;
 854 }
 855 
 856 // Helper for initialization of arrays, creating a ClearArray.
 857 // It writes zero bits in [start..end), within the body of an array object.
 858 // The memory effects are all chained onto the 'adr_type' alias category.
 859 //
 860 // Since the object is otherwise uninitialized, we are free
 861 // to put a little "slop" around the edges of the cleared area,
 862 // as long as it does not go back into the array's header,
 863 // or beyond the array end within the heap.
 864 //
 865 // The lower edge can be rounded down to the nearest jint and the
 866 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
 867 //
 868 // Arguments:
 869 //   adr_type           memory slice where writes are generated
 870 //   dest               oop of the destination array
 871 //   basic_elem_type    element type of the destination
 872 //   slice_idx          array index of first element to store
 873 //   slice_len          number of elements to store (or null)
 874 //   dest_size          total size in bytes of the array object
 875 //
 876 // Exactly one of slice_len or dest_size must be non-null.
 877 // If dest_size is non-null, zeroing extends to the end of the object.
 878 // If slice_len is non-null, the slice_idx value must be a constant.
 879 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
 880                                             const TypePtr* adr_type,
 881                                             Node* dest,


 882                                             BasicType basic_elem_type,
 883                                             Node* slice_idx,
 884                                             Node* slice_len,
 885                                             Node* dest_size) {
 886   // one or the other but not both of slice_len and dest_size:
 887   assert((slice_len != nullptr? 1: 0) + (dest_size != nullptr? 1: 0) == 1, "");
 888   if (slice_len == nullptr)  slice_len = top();
 889   if (dest_size == nullptr)  dest_size = top();
 890 
 891   uint alias_idx = C->get_alias_index(adr_type);
 892 
 893   // operate on this memory slice:
 894   Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
 895 
 896   // scaling and rounding of indexes:
 897   int scale = exact_log2(type2aelembytes(basic_elem_type));
 898   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 899   int clear_low = (-1 << scale) & (BytesPerInt  - 1);
 900   int bump_bit  = (-1 << scale) & BytesPerInt;
 901 
 902   // determine constant starts and ends
 903   const intptr_t BIG_NEG = -128;
 904   assert(BIG_NEG + 2*abase < 0, "neg enough");
 905   intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
 906   intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
 907   if (slice_len_con == 0) {
 908     return;                     // nothing to do here
 909   }
 910   intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
 911   intptr_t end_con   = _igvn.find_intptr_t_con(dest_size, -1);
 912   if (slice_idx_con >= 0 && slice_len_con >= 0) {
 913     assert(end_con < 0, "not two cons");
 914     end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
 915                        BytesPerLong);
 916   }
 917 
 918   if (start_con >= 0 && end_con >= 0) {
 919     // Constant start and end.  Simple.
 920     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 921                                        start_con, end_con, &_igvn);
 922   } else if (start_con >= 0 && dest_size != top()) {
 923     // Constant start, pre-rounded end after the tail of the array.
 924     Node* end = dest_size;
 925     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 926                                        start_con, end, &_igvn);
 927   } else if (start_con >= 0 && slice_len != top()) {
 928     // Constant start, non-constant end.  End needs rounding up.
 929     // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
 930     intptr_t end_base  = abase + (slice_idx_con << scale);
 931     int      end_round = (-1 << scale) & (BytesPerLong  - 1);
 932     Node*    end       = ConvI2X(slice_len);
 933     if (scale != 0)
 934       end = transform_later(new LShiftXNode(end, intcon(scale) ));
 935     end_base += end_round;
 936     end = transform_later(new AddXNode(end, MakeConX(end_base)) );
 937     end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
 938     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 939                                        start_con, end, &_igvn);
 940   } else if (start_con < 0 && dest_size != top()) {
 941     // Non-constant start, pre-rounded end after the tail of the array.
 942     // This is almost certainly a "round-to-end" operation.
 943     Node* start = slice_idx;
 944     start = ConvI2X(start);
 945     if (scale != 0)
 946       start = transform_later(new LShiftXNode( start, intcon(scale) ));
 947     start = transform_later(new AddXNode(start, MakeConX(abase)) );
 948     if ((bump_bit | clear_low) != 0) {
 949       int to_clear = (bump_bit | clear_low);
 950       // Align up mod 8, then store a jint zero unconditionally
 951       // just before the mod-8 boundary.
 952       if (((abase + bump_bit) & ~to_clear) - bump_bit
 953           < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
 954         bump_bit = 0;
 955         assert((abase & to_clear) == 0, "array base must be long-aligned");
 956       } else {
 957         // Bump 'start' up to (or past) the next jint boundary:
 958         start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
 959         assert((abase & clear_low) == 0, "array base must be int-aligned");
 960       }
 961       // Round bumped 'start' down to jlong boundary in body of array.
 962       start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
 963       if (bump_bit != 0) {
 964         // Store a zero to the immediately preceding jint:
 965         Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
 966         Node* p1 = basic_plus_adr(dest, x1);
 967         mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);






 968         mem = transform_later(mem);
 969       }
 970     }
 971     Node* end = dest_size; // pre-rounded
 972     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 973                                        start, end, &_igvn);
 974   } else {
 975     // Non-constant start, unrounded non-constant end.
 976     // (Nobody zeroes a random midsection of an array using this routine.)
 977     ShouldNotReachHere();       // fix caller
 978   }
 979 
 980   // Done.
 981   merge_mem->set_memory_at(alias_idx, mem);
 982 }
 983 
 984 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
 985                                                 const TypePtr* adr_type,
 986                                                 BasicType basic_elem_type,
 987                                                 AllocateNode* alloc,
 988                                                 Node* src,  Node* src_offset,
 989                                                 Node* dest, Node* dest_offset,
 990                                                 Node* dest_size, bool dest_uninitialized) {
 991   // See if there is an advantage from block transfer.
 992   int scale = exact_log2(type2aelembytes(basic_elem_type));
 993   if (scale >= LogBytesPerLong)
 994     return false;               // it is already a block transfer
 995 
 996   // Look at the alignment of the starting offsets.
 997   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 998 
 999   intptr_t src_off_con  = (intptr_t) _igvn.find_int_con(src_offset, -1);
1000   intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
1001   if (src_off_con < 0 || dest_off_con < 0) {
1002     // At present, we can only understand constants.
1003     return false;
1004   }
1005 
1006   intptr_t src_off  = abase + (src_off_con  << scale);
1007   intptr_t dest_off = abase + (dest_off_con << scale);
1008 
1009   if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1010     // Non-aligned; too bad.
1011     // One more chance:  Pick off an initial 32-bit word.
1012     // This is a common case, since abase can be odd mod 8.
1013     if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1014         ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1015       Node* sptr = basic_plus_adr(src,  src_off);
1016       Node* dptr = basic_plus_adr(dest, dest_off);
1017       const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
1018       assert(s_adr_type->isa_aryptr(), "impossible slice");
1019       uint s_alias_idx = C->get_alias_index(s_adr_type);
1020       uint d_alias_idx = C->get_alias_index(adr_type);
1021       bool is_mismatched = (basic_elem_type != T_INT);
1022       Node* sval = transform_later(
1023           LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
1024                          TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
1025                          false /*require_atomic_access*/, false /*unaligned*/, is_mismatched));
1026       Node* st = transform_later(
1027           StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
1028                           sval, T_INT, MemNode::unordered));
1029       if (is_mismatched) {
1030         st->as_Store()->set_mismatched_access();
1031       }
1032       (*mem)->set_memory_at(d_alias_idx, st);
1033       src_off += BytesPerInt;
1034       dest_off += BytesPerInt;
1035     } else {
1036       return false;
1037     }
1038   }
1039   assert(src_off % BytesPerLong == 0, "");
1040   assert(dest_off % BytesPerLong == 0, "");
1041 
1042   // Do this copy by giant steps.
1043   Node* sptr  = basic_plus_adr(src,  src_off);
1044   Node* dptr  = basic_plus_adr(dest, dest_off);
1045   Node* countx = dest_size;
1046   countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
1047   countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
1048 
1049   bool disjoint_bases = true;   // since alloc isn't null
1050   generate_unchecked_arraycopy(ctrl, mem,
1051                                adr_type, T_LONG, disjoint_bases,
1052                                sptr, nullptr, dptr, nullptr, countx, dest_uninitialized);
1053 
1054   return true;
1055 }
1056 
1057 // Helper function; generates code for the slow case.
1058 // We make a call to a runtime method which emulates the native method,
1059 // but without the native wrapper overhead.
1060 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
1061                                                         Node** ctrl, Node* mem, Node** io,
1062                                                         const TypePtr* adr_type,
1063                                                         Node* src,  Node* src_offset,
1064                                                         Node* dest, Node* dest_offset,
1065                                                         Node* copy_length, bool dest_uninitialized) {
1066   assert(!dest_uninitialized, "Invariant");
1067 
1068   const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1069   CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1070                                           "slow_arraycopy", TypePtr::BOTTOM);
1071 
1072   call->init_req(TypeFunc::Control, *ctrl);
1073   call->init_req(TypeFunc::I_O    , *io);
1074   call->init_req(TypeFunc::Memory , mem);
1075   call->init_req(TypeFunc::ReturnAdr, top());
1076   call->init_req(TypeFunc::FramePtr, top());
1077   call->init_req(TypeFunc::Parms+0, src);
1078   call->init_req(TypeFunc::Parms+1, src_offset);
1079   call->init_req(TypeFunc::Parms+2, dest);
1080   call->init_req(TypeFunc::Parms+3, dest_offset);
1081   call->init_req(TypeFunc::Parms+4, copy_length);
1082   call->copy_call_debug_info(&_igvn, ac);
1083 
1084   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1085   _igvn.replace_node(ac, call);
1086   transform_later(call);
1087 
1088   call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1089   *ctrl = _callprojs.fallthrough_catchproj->clone();
1090   transform_later(*ctrl);
1091 
1092   Node* m = _callprojs.fallthrough_memproj->clone();
1093   transform_later(m);
1094 
1095   uint alias_idx = C->get_alias_index(adr_type);
1096   MergeMemNode* out_mem;
1097   if (alias_idx != Compile::AliasIdxBot) {
1098     out_mem = MergeMemNode::make(mem);
1099     out_mem->set_memory_at(alias_idx, m);
1100   } else {
1101     out_mem = MergeMemNode::make(m);
1102   }
1103   transform_later(out_mem);
1104 
1105   // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1106   // could be null. Skip clone and update null fallthrough_ioproj.
1107   if (_callprojs.fallthrough_ioproj != nullptr) {
1108     *io = _callprojs.fallthrough_ioproj->clone();
1109     transform_later(*io);
1110   } else {
1111     *io = nullptr;
1112   }
1113 
1114   return out_mem;
1115 }
1116 
1117 // Helper function; generates code for cases requiring runtime checks.
1118 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1119                                                      const TypePtr* adr_type,
1120                                                      Node* dest_elem_klass,
1121                                                      Node* src,  Node* src_offset,
1122                                                      Node* dest, Node* dest_offset,
1123                                                      Node* copy_length, bool dest_uninitialized) {
1124   if ((*ctrl)->is_top())  return nullptr;
1125 
1126   address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1127   if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1128     return nullptr;
1129   }
1130 
1131   // Pick out the parameters required to perform a store-check
1132   // for the target array.  This is an optimistic check.  It will
1133   // look in each non-null element's class, at the desired klass's
1134   // super_check_offset, for the desired klass.
1135   int sco_offset = in_bytes(Klass::super_check_offset_offset());
1136   Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1137   Node* n3 = new LoadINode(nullptr, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1138   Node* check_offset = ConvI2X(transform_later(n3));
1139   Node* check_value  = dest_elem_klass;
1140 
1141   Node* src_start  = array_element_address(src,  src_offset,  T_OBJECT);
1142   Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1143 
1144   const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1145   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1146                               src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1147 
1148   finish_arraycopy_call(call, ctrl, mem, adr_type);
1149 
1150   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1151   transform_later(proj);
1152 
1153   return proj;
1154 }
1155 
1156 // Helper function; generates code for cases requiring runtime checks.
1157 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1158                                                    const TypePtr* adr_type,
1159                                                    Node* src,  Node* src_offset,
1160                                                    Node* dest, Node* dest_offset,
1161                                                    Node* copy_length, bool dest_uninitialized) {
1162   if ((*ctrl)->is_top()) return nullptr;
1163   assert(!dest_uninitialized, "Invariant");
1164 
1165   address copyfunc_addr = StubRoutines::generic_arraycopy();
1166   if (copyfunc_addr == nullptr) { // Stub was not generated, go slow path.
1167     return nullptr;
1168   }
1169 
1170   const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1171   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1172                               src, src_offset, dest, dest_offset, copy_length);
1173 
1174   finish_arraycopy_call(call, ctrl, mem, adr_type);
1175 
1176   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1177   transform_later(proj);
1178 
1179   return proj;
1180 }
1181 
1182 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1183 void PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1184                                                     const TypePtr* adr_type,
1185                                                     BasicType basic_elem_type,
1186                                                     bool disjoint_bases,
1187                                                     Node* src,  Node* src_offset,
1188                                                     Node* dest, Node* dest_offset,
1189                                                     Node* copy_length, bool dest_uninitialized) {
1190   if ((*ctrl)->is_top()) {
1191     return;
1192   }
1193 
1194   Node* src_start  = src;
1195   Node* dest_start = dest;
1196   if (src_offset != nullptr || dest_offset != nullptr) {
1197     src_start =  array_element_address(src, src_offset, basic_elem_type);
1198     dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1199   }
1200 
1201   // Figure out which arraycopy runtime method to call.
1202   const char* copyfunc_name = "arraycopy";
1203   address     copyfunc_addr =
1204       basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1205                           disjoint_bases, copyfunc_name, dest_uninitialized);
1206 
1207   Node* result_memory = nullptr;
1208   RegionNode* exit_block = nullptr;
1209   if (ArrayOperationPartialInlineSize > 0 && is_subword_type(basic_elem_type) &&
1210     Matcher::vector_width_in_bytes(basic_elem_type) >= 16) {
1211     generate_partial_inlining_block(ctrl, mem, adr_type, &exit_block, &result_memory,
1212                                     copy_length, src_start, dest_start, basic_elem_type);
1213   }
1214 
1215   const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1216   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1217                               src_start, dest_start, copy_length XTOP);
1218 
1219   finish_arraycopy_call(call, ctrl, mem, adr_type);
1220 
1221   // Connecting remaining edges for exit_block coming from stub_block.
1222   if (exit_block) {
1223     exit_block->init_req(2, *ctrl);
1224 
1225     // Memory edge corresponding to stub_region.
1226     result_memory->init_req(2, *mem);
1227 
1228     uint alias_idx = C->get_alias_index(adr_type);
1229     if (alias_idx != Compile::AliasIdxBot) {
1230       *mem = MergeMemNode::make(*mem);
1231       (*mem)->set_memory_at(alias_idx, result_memory);
1232     } else {
1233       *mem = MergeMemNode::make(result_memory);
1234     }
1235     transform_later(*mem);
1236     *ctrl = exit_block;
1237   }
1238 }
1239 




































1240 #undef XTOP
1241 
1242 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1243   Node* ctrl = ac->in(TypeFunc::Control);
1244   Node* io = ac->in(TypeFunc::I_O);
1245   Node* src = ac->in(ArrayCopyNode::Src);
1246   Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1247   Node* dest = ac->in(ArrayCopyNode::Dest);
1248   Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1249   Node* length = ac->in(ArrayCopyNode::Length);
1250   MergeMemNode* merge_mem = nullptr;
1251 
1252   if (ac->is_clonebasic()) {
1253     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1254     bs->clone_at_expansion(this, ac);
1255     return;
1256   } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {
1257     Node* mem = ac->in(TypeFunc::Memory);
1258     merge_mem = MergeMemNode::make(mem);
1259     transform_later(merge_mem);













1260 
1261     AllocateArrayNode* alloc = nullptr;

1262     if (ac->is_alloc_tightly_coupled()) {
1263       alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1264       assert(alloc != nullptr, "expect alloc");

1265     }
1266 
1267     const TypePtr* adr_type = _igvn.type(dest)->is_oopptr()->add_offset(Type::OffsetBot);
1268     if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1269       adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();

















1270     }



1271     generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1272                        adr_type, T_OBJECT,
1273                        src, src_offset, dest, dest_offset, length,

1274                        true, ac->has_negative_length_guard());
1275 
1276     return;
1277   }
1278 
1279   AllocateArrayNode* alloc = nullptr;
1280   if (ac->is_alloc_tightly_coupled()) {
1281     alloc = AllocateArrayNode::Ideal_array_allocation(dest);
1282     assert(alloc != nullptr, "expect alloc");
1283   }
1284 
1285   assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1286 
1287   // Compile time checks.  If any of these checks cannot be verified at compile time,
1288   // we do not make a fast path for this call.  Instead, we let the call remain as it
1289   // is.  The checks we choose to mandate at compile time are:
1290   //
1291   // (1) src and dest are arrays.
1292   const Type* src_type = src->Value(&_igvn);
1293   const Type* dest_type = dest->Value(&_igvn);
1294   const TypeAryPtr* top_src = src_type->isa_aryptr();
1295   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1296 
1297   BasicType src_elem = T_CONFLICT;
1298   BasicType dest_elem = T_CONFLICT;
1299 
1300   if (top_src != nullptr && top_src->elem() != Type::BOTTOM) {
1301     src_elem = top_src->elem()->array_element_basic_type();
1302   }
1303   if (top_dest != nullptr && top_dest->elem() != Type::BOTTOM) {
1304     dest_elem = top_dest->elem()->array_element_basic_type();
1305   }
1306   if (is_reference_type(src_elem, true)) src_elem = T_OBJECT;
1307   if (is_reference_type(dest_elem, true)) dest_elem = T_OBJECT;
1308 
1309   if (ac->is_arraycopy_validated() &&
1310       dest_elem != T_CONFLICT &&
1311       src_elem == T_CONFLICT) {
1312     src_elem = dest_elem;
1313   }
1314 
1315   if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1316     // Conservatively insert a memory barrier on all memory slices.
1317     // Do not let writes into the source float below the arraycopy.
1318     {
1319       Node* mem = ac->in(TypeFunc::Memory);
1320       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder, Compile::AliasIdxBot);
1321 
1322       merge_mem = MergeMemNode::make(mem);
1323       transform_later(merge_mem);
1324     }
1325 
1326     // Call StubRoutines::generic_arraycopy stub.
1327     Node* mem = generate_arraycopy(ac, nullptr, &ctrl, merge_mem, &io,
1328                                    TypeRawPtr::BOTTOM, T_CONFLICT,
1329                                    src, src_offset, dest, dest_offset, length,
1330                                    // If a  negative length guard was generated for the ArrayCopyNode,
1331                                    // the length of the array can never be negative.
1332                                    false, ac->has_negative_length_guard());

1333     return;
1334   }
1335 
1336   assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1337 
1338   // (2) src and dest arrays must have elements of the same BasicType
1339   // Figure out the size and type of the elements we will be copying.
1340   if (src_elem != dest_elem || dest_elem == T_VOID) {







1341     // The component types are not the same or are not recognized.  Punt.
1342     // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1343     {
1344       Node* mem = ac->in(TypeFunc::Memory);
1345       merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1346     }
1347 
1348     _igvn.replace_node(_callprojs.fallthrough_memproj, merge_mem);
1349     if (_callprojs.fallthrough_ioproj != nullptr) {
1350       _igvn.replace_node(_callprojs.fallthrough_ioproj, io);
1351     }
1352     _igvn.replace_node(_callprojs.fallthrough_catchproj, ctrl);
1353     return;
1354   }
1355 
1356   //---------------------------------------------------------------------------
1357   // We will make a fast path for this call to arraycopy.
1358 
1359   // We have the following tests left to perform:
1360   //
1361   // (3) src and dest must not be null.
1362   // (4) src_offset must not be negative.
1363   // (5) dest_offset must not be negative.
1364   // (6) length must not be negative.
1365   // (7) src_offset + length must not exceed length of src.
1366   // (8) dest_offset + length must not exceed length of dest.
1367   // (9) each element of an oop array must be assignable
1368 
1369   {
1370     Node* mem = ac->in(TypeFunc::Memory);
1371     merge_mem = MergeMemNode::make(mem);
1372     transform_later(merge_mem);





1373   }


1374 
1375   RegionNode* slow_region = new RegionNode(1);
1376   transform_later(slow_region);
1377 
1378   if (!ac->is_arraycopy_validated()) {
1379     // (3) operands must not be null
1380     // We currently perform our null checks with the null_check routine.
1381     // This means that the null exceptions will be reported in the caller
1382     // rather than (correctly) reported inside of the native arraycopy call.
1383     // This should be corrected, given time.  We do our null check with the
1384     // stack pointer restored.
1385     // null checks done library_call.cpp
1386 
1387     // (4) src_offset must not be negative.
1388     generate_negative_guard(&ctrl, src_offset, slow_region);
1389 
1390     // (5) dest_offset must not be negative.
1391     generate_negative_guard(&ctrl, dest_offset, slow_region);
1392 
1393     // (6) length must not be negative (moved to generate_arraycopy()).
1394     // generate_negative_guard(length, slow_region);
1395 
1396     // (7) src_offset + length must not exceed length of src.
1397     Node* alen = ac->in(ArrayCopyNode::SrcLen);
1398     assert(alen != nullptr, "need src len");
1399     generate_limit_guard(&ctrl,
1400                          src_offset, length,
1401                          alen,
1402                          slow_region);
1403 
1404     // (8) dest_offset + length must not exceed length of dest.
1405     alen = ac->in(ArrayCopyNode::DestLen);
1406     assert(alen != nullptr, "need dest len");
1407     generate_limit_guard(&ctrl,
1408                          dest_offset, length,
1409                          alen,
1410                          slow_region);
1411 
1412     // (9) each element of an oop array must be assignable
1413     // The generate_arraycopy subroutine checks this.








1414   }

1415   // This is where the memory effects are placed:
1416   const TypePtr* adr_type = nullptr;
1417   if (ac->_dest_type != TypeOopPtr::BOTTOM) {





1418     adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1419   } else {
1420     adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1421   }
1422 
1423   generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1424                      adr_type, dest_elem,
1425                      src, src_offset, dest, dest_offset, length,

1426                      // If a  negative length guard was generated for the ArrayCopyNode,
1427                      // the length of the array can never be negative.
1428                      false, ac->has_negative_length_guard(), slow_region);

1429 }
--- EOF ---