1 /*
   2  * Copyright (c) 2012, 2021, 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 "precompiled.hpp"

  26 #include "gc/shared/barrierSet.hpp"
  27 #include "gc/shared/tlab_globals.hpp"
  28 #include "opto/arraycopynode.hpp"
  29 #include "oops/objArrayKlass.hpp"
  30 #include "opto/convertnode.hpp"
  31 #include "opto/vectornode.hpp"
  32 #include "opto/graphKit.hpp"
  33 #include "opto/macro.hpp"
  34 #include "opto/runtime.hpp"
  35 #include "opto/castnode.hpp"
  36 #include "runtime/stubRoutines.hpp"
  37 #include "utilities/align.hpp"
  38 #include "utilities/powerOfTwo.hpp"
  39 
  40 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, Node* precedent) {
  41   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
  42   mb->init_req(TypeFunc::Control, *ctrl);
  43   mb->init_req(TypeFunc::Memory, *mem);
  44   transform_later(mb);
  45   *ctrl = new ProjNode(mb,TypeFunc::Control);
  46   transform_later(*ctrl);
  47   Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
  48   transform_later(mem_proj);
  49   *mem = mem_proj;
  50 }
  51 
  52 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
  53   uint shift  = exact_log2(type2aelembytes(elembt));
  54   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
  55   Node* base =  basic_plus_adr(ary, header);
  56 #ifdef _LP64
  57   // see comment in GraphKit::array_element_address
  58   int index_max = max_jint - 1;  // array size is max_jint, index is one less
  59   const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
  60   idx = transform_later( new ConvI2LNode(idx, lidxtype) );
  61 #endif
  62   Node* scale = new LShiftXNode(idx, intcon(shift));
  63   transform_later(scale);
  64   return basic_plus_adr(ary, base, scale);
  65 }
  66 
  67 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
  68   return transform_later(new ConvI2LNode(offset));
  69 }
  70 
  71 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
  72                                        const TypeFunc* call_type, address call_addr,
  73                                        const char* call_name,
  74                                        const TypePtr* adr_type,
  75                                        Node* parm0, Node* parm1,
  76                                        Node* parm2, Node* parm3,
  77                                        Node* parm4, Node* parm5,
  78                                        Node* parm6, Node* parm7) {
  79   Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
  80   call->init_req(TypeFunc::Control, ctrl);
  81   call->init_req(TypeFunc::I_O    , top());
  82   call->init_req(TypeFunc::Memory , mem);
  83   call->init_req(TypeFunc::ReturnAdr, top());
  84   call->init_req(TypeFunc::FramePtr, top());
  85 
  86   // Hook each parm in order.  Stop looking at the first NULL.
  87   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
  88   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
  89   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
  90   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
  91   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
  92   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
  93   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
  94   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
  95     /* close each nested if ===> */  } } } } } } } }
  96   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
  97 
  98   return call;
  99 }
 100 
 101 
 102 //------------------------------generate_guard---------------------------
 103 // Helper function for generating guarded fast-slow graph structures.
 104 // The given 'test', if true, guards a slow path.  If the test fails
 105 // then a fast path can be taken.  (We generally hope it fails.)
 106 // In all cases, GraphKit::control() is updated to the fast path.
 107 // The returned value represents the control for the slow path.
 108 // The return value is never 'top'; it is either a valid control
 109 // or NULL if it is obvious that the slow path can never be taken.
 110 // Also, if region and the slow control are not NULL, the slow edge
 111 // is appended to the region.
 112 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
 113   if ((*ctrl)->is_top()) {
 114     // Already short circuited.
 115     return NULL;
 116   }
 117   // Build an if node and its projections.
 118   // If test is true we take the slow path, which we assume is uncommon.
 119   if (_igvn.type(test) == TypeInt::ZERO) {
 120     // The slow branch is never taken.  No need to build this guard.
 121     return NULL;
 122   }
 123 
 124   IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
 125   transform_later(iff);
 126 
 127   Node* if_slow = new IfTrueNode(iff);
 128   transform_later(if_slow);
 129 
 130   if (region != NULL) {
 131     region->add_req(if_slow);
 132   }
 133 
 134   Node* if_fast = new IfFalseNode(iff);
 135   transform_later(if_fast);
 136 
 137   *ctrl = if_fast;
 138 
 139   return if_slow;
 140 }
 141 
 142 inline Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
 143   return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
 144 }
 145 




 146 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
 147   if ((*ctrl)->is_top())
 148     return;                // already stopped
 149   if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
 150     return;                // index is already adequately typed
 151   Node* cmp_lt = new CmpINode(index, intcon(0));
 152   transform_later(cmp_lt);
 153   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 154   transform_later(bol_lt);
 155   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 156 }
 157 
 158 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
 159   if ((*ctrl)->is_top())
 160     return;                // already stopped
 161   bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
 162   if (zero_offset && subseq_length->eqv_uncast(array_length))
 163     return;                // common case of whole-array copy
 164   Node* last = subseq_length;
 165   if (!zero_offset) {            // last += offset
 166     last = new AddINode(last, offset);
 167     transform_later(last);
 168   }
 169   Node* cmp_lt = new CmpUNode(array_length, last);
 170   transform_later(cmp_lt);
 171   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 172   transform_later(bol_lt);
 173   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 174 }
 175 
 176 //
 177 // Partial in-lining handling for smaller conjoint/disjoint array copies having
 178 // length(in bytes) less than ArrayOperationPartialInlineSize.
 179 //  if (length <= ArrayOperationPartialInlineSize) {
 180 //    partial_inlining_block:
 181 //      mask = Mask_Gen
 182 //      vload = LoadVectorMasked src , mask
 183 //      StoreVectorMasked dst, mask, vload
 184 //  } else {
 185 //    stub_block:
 186 //      callstub array_copy
 187 //  }
 188 //  exit_block:
 189 //    Phi = label partial_inlining_block:mem , label stub_block:mem (filled by caller)
 190 //    mem = MergeMem (Phi)
 191 //    control = stub_block
 192 //
 193 //  Exit_block and associated phi(memory) are partially initialized for partial_in-lining_block
 194 //  edges. Remaining edges for exit_block coming from stub_block are connected by the caller
 195 //  post stub nodes creation.
 196 //
 197 
 198 void PhaseMacroExpand::generate_partial_inlining_block(Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type,
 199                                                        RegionNode** exit_block, Node** result_memory, Node* length,
 200                                                        Node* src_start, Node* dst_start, BasicType type) {
 201   const TypePtr *src_adr_type = _igvn.type(src_start)->isa_ptr();
 202   Node* inline_block = NULL;
 203   Node* stub_block = NULL;
 204 
 205   int const_len = -1;
 206   const TypeInt* lty = NULL;
 207   uint shift  = exact_log2(type2aelembytes(type));
 208   if (length->Opcode() == Op_ConvI2L) {
 209     lty = _igvn.type(length->in(1))->isa_int();
 210   } else  {
 211     lty = _igvn.type(length)->isa_int();
 212   }
 213   if (lty && lty->is_con()) {
 214     const_len = lty->get_con() << shift;
 215   }
 216 
 217   // Return if copy length is greater than partial inline size limit or
 218   // target does not supports masked load/stores.
 219   int lane_count = ArrayCopyNode::get_partial_inline_vector_lane_count(type, const_len);
 220   if ( const_len > ArrayOperationPartialInlineSize ||
 221       !Matcher::match_rule_supported_vector(Op_LoadVectorMasked, lane_count, type)  ||
 222       !Matcher::match_rule_supported_vector(Op_StoreVectorMasked, lane_count, type) ||
 223       !Matcher::match_rule_supported_vector(Op_VectorMaskGen, lane_count, type)) {
 224     return;
 225   }
 226 
 227   int inline_limit = ArrayOperationPartialInlineSize / type2aelembytes(type);
 228   Node* casted_length = new CastLLNode(*ctrl, length, TypeLong::make(0, inline_limit, Type::WidenMin));
 229   transform_later(casted_length);
 230   Node* copy_bytes = new LShiftXNode(length, intcon(shift));
 231   transform_later(copy_bytes);
 232 
 233   Node* cmp_le = new CmpULNode(copy_bytes, longcon(ArrayOperationPartialInlineSize));
 234   transform_later(cmp_le);
 235   Node* bol_le = new BoolNode(cmp_le, BoolTest::le);
 236   transform_later(bol_le);
 237   inline_block  = generate_guard(ctrl, bol_le, NULL, PROB_FAIR);
 238   stub_block = *ctrl;
 239 
 240   Node* mask_gen =  new VectorMaskGenNode(casted_length, TypeVect::VECTMASK, type);
 241   transform_later(mask_gen);
 242 
 243   unsigned vec_size = lane_count *  type2aelembytes(type);
 244   if (C->max_vector_size() < vec_size) {
 245     C->set_max_vector_size(vec_size);
 246   }
 247 
 248   const TypeVect * vt = TypeVect::make(type, lane_count);
 249   Node* mm = (*mem)->memory_at(C->get_alias_index(src_adr_type));
 250   Node* masked_load = new LoadVectorMaskedNode(inline_block, mm, src_start,
 251                                                src_adr_type, vt, mask_gen);
 252   transform_later(masked_load);
 253 
 254   mm = (*mem)->memory_at(C->get_alias_index(adr_type));
 255   Node* masked_store = new StoreVectorMaskedNode(inline_block, mm, dst_start,
 256                                                  masked_load, adr_type, mask_gen);
 257   transform_later(masked_store);
 258 
 259   // Convergence region for inline_block and stub_block.
 260   *exit_block = new RegionNode(3);
 261   transform_later(*exit_block);
 262   (*exit_block)->init_req(1, inline_block);
 263   *result_memory = new PhiNode(*exit_block, Type::MEMORY, adr_type);
 264   transform_later(*result_memory);
 265   (*result_memory)->init_req(1, masked_store);
 266 
 267   *ctrl = stub_block;
 268 }
 269 
 270 
 271 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
 272   if ((*ctrl)->is_top())  return NULL;
 273 
 274   if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
 275     return NULL;                // index is already adequately typed
 276   Node* cmp_le = new CmpINode(index, intcon(0));
 277   transform_later(cmp_le);
 278   BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
 279   Node* bol_le = new BoolNode(cmp_le, le_or_eq);
 280   transform_later(bol_le);
 281   Node* is_notp = generate_guard(ctrl, bol_le, NULL, PROB_MIN);
 282 
 283   return is_notp;
 284 }
 285 




















 286 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
 287   transform_later(call);
 288 
 289   *ctrl = new ProjNode(call,TypeFunc::Control);
 290   transform_later(*ctrl);
 291   Node* newmem = new ProjNode(call, TypeFunc::Memory);
 292   transform_later(newmem);
 293 
 294   uint alias_idx = C->get_alias_index(adr_type);
 295   if (alias_idx != Compile::AliasIdxBot) {
 296     *mem = MergeMemNode::make(*mem);
 297     (*mem)->set_memory_at(alias_idx, newmem);
 298   } else {
 299     *mem = MergeMemNode::make(newmem);
 300   }
 301   transform_later(*mem);
 302 }
 303 
 304 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
 305                                               Node* src_offset,
 306                                               Node* dest_offset,
 307                                               bool disjoint_bases,
 308                                               const char* &name,
 309                                               bool dest_uninitialized) {
 310   const TypeInt* src_offset_inttype  = _igvn.find_int_type(src_offset);
 311   const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);
 312 
 313   bool aligned = false;
 314   bool disjoint = disjoint_bases;
 315 
 316   // if the offsets are the same, we can treat the memory regions as
 317   // disjoint, because either the memory regions are in different arrays,
 318   // or they are identical (which we can treat as disjoint.)  We can also
 319   // treat a copy with a destination index  less that the source index
 320   // as disjoint since a low->high copy will work correctly in this case.
 321   if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
 322       dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
 323     // both indices are constants
 324     int s_offs = src_offset_inttype->get_con();
 325     int d_offs = dest_offset_inttype->get_con();
 326     int element_size = type2aelembytes(t);
 327     aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
 328               ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
 329     if (s_offs >= d_offs)  disjoint = true;
 330   } else if (src_offset == dest_offset && src_offset != NULL) {
 331     // This can occur if the offsets are identical non-constants.
 332     disjoint = true;
 333   }
 334 
 335   return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
 336 }
 337 



















 338 #define XTOP LP64_ONLY(COMMA top())
 339 
 340 // Generate an optimized call to arraycopy.
 341 // Caller must guard against non-arrays.
 342 // Caller must determine a common array basic-type for both arrays.
 343 // Caller must validate offsets against array bounds.
 344 // The slow_region has already collected guard failure paths
 345 // (such as out of bounds length or non-conformable array types).
 346 // The generated code has this shape, in general:
 347 //
 348 //     if (length == 0)  return   // via zero_path
 349 //     slowval = -1
 350 //     if (types unknown) {
 351 //       slowval = call generic copy loop
 352 //       if (slowval == 0)  return  // via checked_path
 353 //     } else if (indexes in bounds) {
 354 //       if ((is object array) && !(array type check)) {
 355 //         slowval = call checked copy loop
 356 //         if (slowval == 0)  return  // via checked_path
 357 //       } else {
 358 //         call bulk copy loop
 359 //         return  // via fast_path
 360 //       }
 361 //     }
 362 //     // adjust params for remaining work:
 363 //     if (slowval != -1) {
 364 //       n = -1^slowval; src_offset += n; dest_offset += n; length -= n
 365 //     }
 366 //   slow_region:
 367 //     call slow arraycopy(src, src_offset, dest, dest_offset, length)
 368 //     return  // via slow_call_path
 369 //
 370 // This routine is used from several intrinsics:  System.arraycopy,
 371 // Object.clone (the array subcase), and Arrays.copyOf[Range].
 372 //
 373 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
 374                                            Node** ctrl, MergeMemNode* mem, Node** io,
 375                                            const TypePtr* adr_type,
 376                                            BasicType basic_elem_type,
 377                                            Node* src,  Node* src_offset,
 378                                            Node* dest, Node* dest_offset,
 379                                            Node* copy_length,

 380                                            bool disjoint_bases,
 381                                            bool length_never_negative,
 382                                            RegionNode* slow_region) {
 383   if (slow_region == NULL) {
 384     slow_region = new RegionNode(1);
 385     transform_later(slow_region);
 386   }
 387 
 388   Node* original_dest = dest;
 389   bool  dest_needs_zeroing   = false;
 390   bool  acopy_to_uninitialized = false;


 391 
 392   // See if this is the initialization of a newly-allocated array.
 393   // If so, we will take responsibility here for initializing it to zero.
 394   // (Note:  Because tightly_coupled_allocation performs checks on the
 395   // out-edges of the dest, we need to avoid making derived pointers
 396   // from it until we have checked its uses.)
 397   if (ReduceBulkZeroing
 398       && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
 399       && basic_elem_type != T_CONFLICT // avoid corner case
 400       && !src->eqv_uncast(dest)
 401       && alloc != NULL
 402       && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0) {
 403     assert(ac->is_alloc_tightly_coupled(), "sanity");
 404     // acopy to uninitialized tightly coupled allocations
 405     // needs zeroing outside the copy range
 406     // and the acopy itself will be to uninitialized memory
 407     acopy_to_uninitialized = true;
 408     if (alloc->maybe_set_complete(&_igvn)) {
 409       // "You break it, you buy it."
 410       InitializeNode* init = alloc->initialization();
 411       assert(init->is_complete(), "we just did this");
 412       init->set_complete_with_arraycopy();
 413       assert(dest->is_CheckCastPP(), "sanity");
 414       assert(dest->in(0)->in(0) == init, "dest pinned");
 415       adr_type = TypeRawPtr::BOTTOM;  // all initializations are into raw memory
 416       // From this point on, every exit path is responsible for
 417       // initializing any non-copied parts of the object to zero.
 418       // Also, if this flag is set we make sure that arraycopy interacts properly
 419       // with G1, eliding pre-barriers. See CR 6627983.
 420       dest_needs_zeroing = true;


 421     } else {
 422       // dest_need_zeroing = false;
 423     }
 424   } else {
 425     // No zeroing elimination needed here.
 426     alloc                  = NULL;
 427     acopy_to_uninitialized = false;
 428     //original_dest        = dest;
 429     //dest_needs_zeroing   = false;
 430   }
 431 
 432   uint alias_idx = C->get_alias_index(adr_type);
 433 
 434   // Results are placed here:
 435   enum { fast_path        = 1,  // normal void-returning assembly stub
 436          checked_path     = 2,  // special assembly stub with cleanup
 437          slow_call_path   = 3,  // something went wrong; call the VM
 438          zero_path        = 4,  // bypass when length of copy is zero
 439          bcopy_path       = 5,  // copy primitive array by 64-bit blocks
 440          PATH_LIMIT       = 6
 441   };
 442   RegionNode* result_region = new RegionNode(PATH_LIMIT);
 443   PhiNode*    result_i_o    = new PhiNode(result_region, Type::ABIO);
 444   PhiNode*    result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
 445   assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
 446   transform_later(result_region);
 447   transform_later(result_i_o);
 448   transform_later(result_memory);
 449 
 450   // The slow_control path:
 451   Node* slow_control;
 452   Node* slow_i_o = *io;
 453   Node* slow_mem = mem->memory_at(alias_idx);
 454   DEBUG_ONLY(slow_control = (Node*) badAddress);
 455 
 456   // Checked control path:
 457   Node* checked_control = top();
 458   Node* checked_mem     = NULL;
 459   Node* checked_i_o     = NULL;
 460   Node* checked_value   = NULL;
 461 
 462   if (basic_elem_type == T_CONFLICT) {
 463     assert(!dest_needs_zeroing, "");
 464     Node* cv = generate_generic_arraycopy(ctrl, &mem,
 465                                           adr_type,
 466                                           src, src_offset, dest, dest_offset,
 467                                           copy_length, acopy_to_uninitialized);
 468     if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
 469     checked_control = *ctrl;
 470     checked_i_o     = *io;
 471     checked_mem     = mem->memory_at(alias_idx);
 472     checked_value   = cv;
 473     *ctrl = top();
 474   }
 475 
 476   Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
 477   if (not_pos != NULL) {
 478     Node* local_ctrl = not_pos, *local_io = *io;
 479     MergeMemNode* local_mem = MergeMemNode::make(mem);
 480     transform_later(local_mem);
 481 
 482     // (6) length must not be negative.
 483     if (!length_never_negative) {
 484       generate_negative_guard(&local_ctrl, copy_length, slow_region);
 485     }
 486 
 487     // copy_length is 0.
 488     if (dest_needs_zeroing) {
 489       assert(!local_ctrl->is_top(), "no ctrl?");
 490       Node* dest_length = alloc->in(AllocateNode::ALength);
 491       if (copy_length->eqv_uncast(dest_length)
 492           || _igvn.find_int_con(dest_length, 1) <= 0) {
 493         // There is no zeroing to do. No need for a secondary raw memory barrier.
 494       } else {
 495         // Clear the whole thing since there are no source elements to copy.
 496         generate_clear_array(local_ctrl, local_mem,
 497                              adr_type, dest, basic_elem_type,


 498                              intcon(0), NULL,
 499                              alloc->in(AllocateNode::AllocSize));
 500         // Use a secondary InitializeNode as raw memory barrier.
 501         // Currently it is needed only on this path since other
 502         // paths have stub or runtime calls as raw memory barriers.
 503         MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
 504                                           Compile::AliasIdxRaw,
 505                                           top());
 506         transform_later(mb);
 507         mb->set_req(TypeFunc::Control,local_ctrl);
 508         mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
 509         local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
 510         local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
 511 
 512         InitializeNode* init = mb->as_Initialize();
 513         init->set_complete(&_igvn);  // (there is no corresponding AllocateNode)
 514       }
 515     }
 516 
 517     // Present the results of the fast call.
 518     result_region->init_req(zero_path, local_ctrl);
 519     result_i_o   ->init_req(zero_path, local_io);
 520     result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
 521   }
 522 
 523   if (!(*ctrl)->is_top() && dest_needs_zeroing) {
 524     // We have to initialize the *uncopied* part of the array to zero.
 525     // The copy destination is the slice dest[off..off+len].  The other slices
 526     // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
 527     Node* dest_size   = alloc->in(AllocateNode::AllocSize);
 528     Node* dest_length = alloc->in(AllocateNode::ALength);
 529     Node* dest_tail   = transform_later( new AddINode(dest_offset, copy_length));
 530 
 531     // If there is a head section that needs zeroing, do it now.
 532     if (_igvn.find_int_con(dest_offset, -1) != 0) {
 533       generate_clear_array(*ctrl, mem,
 534                            adr_type, dest, basic_elem_type,


 535                            intcon(0), dest_offset,
 536                            NULL);
 537     }
 538 
 539     // Next, perform a dynamic check on the tail length.
 540     // It is often zero, and we can win big if we prove this.
 541     // There are two wins:  Avoid generating the ClearArray
 542     // with its attendant messy index arithmetic, and upgrade
 543     // the copy to a more hardware-friendly word size of 64 bits.
 544     Node* tail_ctl = NULL;
 545     if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
 546       Node* cmp_lt   = transform_later( new CmpINode(dest_tail, dest_length) );
 547       Node* bol_lt   = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
 548       tail_ctl = generate_slow_guard(ctrl, bol_lt, NULL);
 549       assert(tail_ctl != NULL || !(*ctrl)->is_top(), "must be an outcome");
 550     }
 551 
 552     // At this point, let's assume there is no tail.
 553     if (!(*ctrl)->is_top() && alloc != NULL && basic_elem_type != T_OBJECT) {
 554       // There is no tail.  Try an upgrade to a 64-bit copy.
 555       bool didit = false;
 556       {
 557         Node* local_ctrl = *ctrl, *local_io = *io;
 558         MergeMemNode* local_mem = MergeMemNode::make(mem);
 559         transform_later(local_mem);
 560 
 561         didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
 562                                          adr_type, basic_elem_type, alloc,
 563                                          src, src_offset, dest, dest_offset,
 564                                          dest_size, acopy_to_uninitialized);
 565         if (didit) {
 566           // Present the results of the block-copying fast call.
 567           result_region->init_req(bcopy_path, local_ctrl);
 568           result_i_o   ->init_req(bcopy_path, local_io);
 569           result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
 570         }
 571       }
 572       if (didit) {
 573         *ctrl = top();     // no regular fast path
 574       }
 575     }
 576 
 577     // Clear the tail, if any.
 578     if (tail_ctl != NULL) {
 579       Node* notail_ctl = (*ctrl)->is_top() ? NULL : *ctrl;
 580       *ctrl = tail_ctl;
 581       if (notail_ctl == NULL) {
 582         generate_clear_array(*ctrl, mem,
 583                              adr_type, dest, basic_elem_type,


 584                              dest_tail, NULL,
 585                              dest_size);
 586       } else {
 587         // Make a local merge.
 588         Node* done_ctl = transform_later(new RegionNode(3));
 589         Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
 590         done_ctl->init_req(1, notail_ctl);
 591         done_mem->init_req(1, mem->memory_at(alias_idx));
 592         generate_clear_array(*ctrl, mem,
 593                              adr_type, dest, basic_elem_type,


 594                              dest_tail, NULL,
 595                              dest_size);
 596         done_ctl->init_req(2, *ctrl);
 597         done_mem->init_req(2, mem->memory_at(alias_idx));
 598         *ctrl = done_ctl;
 599         mem->set_memory_at(alias_idx, done_mem);
 600       }
 601     }
 602   }
 603 
 604   BasicType copy_type = basic_elem_type;
 605   assert(basic_elem_type != T_ARRAY, "caller must fix this");
 606   if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
 607     // If src and dest have compatible element types, we can copy bits.
 608     // Types S[] and D[] are compatible if D is a supertype of S.
 609     //
 610     // If they are not, we will use checked_oop_disjoint_arraycopy,
 611     // which performs a fast optimistic per-oop check, and backs off
 612     // further to JVM_ArrayCopy on the first per-oop check that fails.
 613     // (Actually, we don't move raw bits only; the GC requires card marks.)
 614 
 615     // We don't need a subtype check for validated copies and Object[].clone()
 616     bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
 617                               ac->is_copyofrange_validated() || ac->is_clone_oop_array();
 618     if (!skip_subtype_check) {
 619       // Get the klass* for both src and dest
 620       Node* src_klass  = ac->in(ArrayCopyNode::SrcKlass);
 621       Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
 622 
 623       assert(src_klass != NULL && dest_klass != NULL, "should have klasses");
 624 
 625       // Generate the subtype check.
 626       // This might fold up statically, or then again it might not.
 627       //
 628       // Non-static example:  Copying List<String>.elements to a new String[].
 629       // The backing store for a List<String> is always an Object[],
 630       // but its elements are always type String, if the generic types
 631       // are correct at the source level.
 632       //
 633       // Test S[] against D[], not S against D, because (probably)
 634       // the secondary supertype cache is less busy for S[] than S.
 635       // This usually only matters when D is an interface.
 636       Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, _igvn);
 637       // Plug failing path into checked_oop_disjoint_arraycopy
 638       if (not_subtype_ctrl != top()) {
 639         Node* local_ctrl = not_subtype_ctrl;
 640         MergeMemNode* local_mem = MergeMemNode::make(mem);
 641         transform_later(local_mem);
 642 
 643         // (At this point we can assume disjoint_bases, since types differ.)
 644         int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
 645         Node* p1 = basic_plus_adr(dest_klass, ek_offset);
 646         Node* n1 = LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
 647         Node* dest_elem_klass = transform_later(n1);
 648         Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
 649                                                 adr_type,
 650                                                 dest_elem_klass,
 651                                                 src, src_offset, dest, dest_offset,
 652                                                 ConvI2X(copy_length), acopy_to_uninitialized);
 653         if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
 654         checked_control = local_ctrl;
 655         checked_i_o     = *io;
 656         checked_mem     = local_mem->memory_at(alias_idx);
 657         checked_value   = cv;
 658       }
 659     }
 660     // At this point we know we do not need type checks on oop stores.
 661 
 662     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 663     if (!bs->array_copy_requires_gc_barriers(alloc != NULL, copy_type, false, false, BarrierSetC2::Expansion)) {
 664       // If we do not need gc barriers, copy using the jint or jlong stub.
 665       copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
 666       assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
 667              "sizes agree");
 668     }
 669   }
 670 
 671   bool is_partial_array_copy = false;
 672   if (!(*ctrl)->is_top()) {
 673     // Generate the fast path, if possible.
 674     Node* local_ctrl = *ctrl;
 675     MergeMemNode* local_mem = MergeMemNode::make(mem);
 676     transform_later(local_mem);
 677     is_partial_array_copy = generate_unchecked_arraycopy(&local_ctrl, &local_mem,
 678                                                          adr_type, copy_type, disjoint_bases,
 679                                                          src, src_offset, dest, dest_offset,
 680                                                          ConvI2X(copy_length), acopy_to_uninitialized);
 681 
 682     // Present the results of the fast call.
 683     result_region->init_req(fast_path, local_ctrl);
 684     result_i_o   ->init_req(fast_path, *io);
 685     result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
 686   }
 687 
 688   // Here are all the slow paths up to this point, in one bundle:
 689   assert(slow_region != NULL, "allocated on entry");
 690   slow_control = slow_region;
 691   DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
 692 
 693   *ctrl = checked_control;
 694   if (!(*ctrl)->is_top()) {
 695     // Clean up after the checked call.
 696     // The returned value is either 0 or -1^K,
 697     // where K = number of partially transferred array elements.
 698     Node* cmp = new CmpINode(checked_value, intcon(0));
 699     transform_later(cmp);
 700     Node* bol = new BoolNode(cmp, BoolTest::eq);
 701     transform_later(bol);
 702     IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
 703     transform_later(iff);
 704 
 705     // If it is 0, we are done, so transfer to the end.
 706     Node* checks_done = new IfTrueNode(iff);
 707     transform_later(checks_done);
 708     result_region->init_req(checked_path, checks_done);
 709     result_i_o   ->init_req(checked_path, checked_i_o);
 710     result_memory->init_req(checked_path, checked_mem);
 711 
 712     // If it is not zero, merge into the slow call.
 713     *ctrl = new IfFalseNode(iff);
 714     transform_later(*ctrl);
 715     RegionNode* slow_reg2 = new RegionNode(3);
 716     PhiNode*    slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
 717     PhiNode*    slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
 718     transform_later(slow_reg2);
 719     transform_later(slow_i_o2);
 720     transform_later(slow_mem2);
 721     slow_reg2  ->init_req(1, slow_control);
 722     slow_i_o2  ->init_req(1, slow_i_o);
 723     slow_mem2  ->init_req(1, slow_mem);
 724     slow_reg2  ->init_req(2, *ctrl);
 725     slow_i_o2  ->init_req(2, checked_i_o);
 726     slow_mem2  ->init_req(2, checked_mem);
 727 
 728     slow_control = slow_reg2;
 729     slow_i_o     = slow_i_o2;
 730     slow_mem     = slow_mem2;
 731 
 732     if (alloc != NULL) {
 733       // We'll restart from the very beginning, after zeroing the whole thing.
 734       // This can cause double writes, but that's OK since dest is brand new.
 735       // So we ignore the low 31 bits of the value returned from the stub.
 736     } else {
 737       // We must continue the copy exactly where it failed, or else
 738       // another thread might see the wrong number of writes to dest.
 739       Node* checked_offset = new XorINode(checked_value, intcon(-1));
 740       Node* slow_offset    = new PhiNode(slow_reg2, TypeInt::INT);
 741       transform_later(checked_offset);
 742       transform_later(slow_offset);
 743       slow_offset->init_req(1, intcon(0));
 744       slow_offset->init_req(2, checked_offset);
 745 
 746       // Adjust the arguments by the conditionally incoming offset.
 747       Node* src_off_plus  = new AddINode(src_offset,  slow_offset);
 748       transform_later(src_off_plus);
 749       Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
 750       transform_later(dest_off_plus);
 751       Node* length_minus  = new SubINode(copy_length, slow_offset);
 752       transform_later(length_minus);
 753 
 754       // Tweak the node variables to adjust the code produced below:
 755       src_offset  = src_off_plus;
 756       dest_offset = dest_off_plus;
 757       copy_length = length_minus;
 758     }
 759   }
 760   *ctrl = slow_control;
 761   if (!(*ctrl)->is_top()) {
 762     Node* local_ctrl = *ctrl, *local_io = slow_i_o;
 763     MergeMemNode* local_mem = MergeMemNode::make(mem);
 764     transform_later(local_mem);
 765 
 766     // Generate the slow path, if needed.
 767     local_mem->set_memory_at(alias_idx, slow_mem);
 768 
 769     if (dest_needs_zeroing) {
 770       generate_clear_array(local_ctrl, local_mem,
 771                            adr_type, dest, basic_elem_type,


 772                            intcon(0), NULL,
 773                            alloc->in(AllocateNode::AllocSize));
 774     }
 775 
 776     local_mem = generate_slow_arraycopy(ac,
 777                                         &local_ctrl, local_mem, &local_io,
 778                                         adr_type,
 779                                         src, src_offset, dest, dest_offset,
 780                                         copy_length, /*dest_uninitialized*/false);
 781 
 782     result_region->init_req(slow_call_path, local_ctrl);
 783     result_i_o   ->init_req(slow_call_path, local_io);
 784     result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
 785   } else {
 786     ShouldNotReachHere(); // no call to generate_slow_arraycopy:
 787                           // projections were not extracted
 788   }
 789 
 790   // Remove unused edges.
 791   for (uint i = 1; i < result_region->req(); i++) {
 792     if (result_region->in(i) == NULL) {
 793       result_region->init_req(i, top());
 794     }
 795   }
 796 
 797   // Finished; return the combined state.
 798   *ctrl = result_region;
 799   *io = result_i_o;
 800   mem->set_memory_at(alias_idx, result_memory);
 801 
 802   // mem no longer guaranteed to stay a MergeMemNode
 803   Node* out_mem = mem;
 804   DEBUG_ONLY(mem = NULL);
 805 
 806   // The memory edges above are precise in order to model effects around
 807   // array copies accurately to allow value numbering of field loads around
 808   // arraycopy.  Such field loads, both before and after, are common in Java
 809   // collections and similar classes involving header/array data structures.
 810   //
 811   // But with low number of register or when some registers are used or killed
 812   // by arraycopy calls it causes registers spilling on stack. See 6544710.
 813   // The next memory barrier is added to avoid it. If the arraycopy can be
 814   // optimized away (which it can, sometimes) then we can manually remove
 815   // the membar also.
 816   //
 817   // Do not let reads from the cloned object float above the arraycopy.
 818   if (alloc != NULL && !alloc->initialization()->does_not_escape()) {
 819     // Do not let stores that initialize this object be reordered with
 820     // a subsequent store that would make this object accessible by
 821     // other threads.
 822     insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore);
 823   } else {




 824     insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
 825   }
 826 
 827   if (is_partial_array_copy) {
 828     assert((*ctrl)->is_Proj(), "MemBar control projection");
 829     assert((*ctrl)->in(0)->isa_MemBar(), "MemBar node");
 830     (*ctrl)->in(0)->isa_MemBar()->set_trailing_partial_array_copy();
 831   }
 832 
 833   _igvn.replace_node(_callprojs.fallthrough_memproj, out_mem);
 834   if (_callprojs.fallthrough_ioproj != NULL) {
 835     _igvn.replace_node(_callprojs.fallthrough_ioproj, *io);
 836   }
 837   _igvn.replace_node(_callprojs.fallthrough_catchproj, *ctrl);
 838 
 839 #ifdef ASSERT
 840   const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
 841   if (dest_t->is_known_instance() && !is_partial_array_copy) {
 842     ArrayCopyNode* ac = NULL;
 843     assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
 844     assert(ac == NULL, "no arraycopy anymore");
 845   }
 846 #endif
 847 
 848   return out_mem;
 849 }
 850 
 851 // Helper for initialization of arrays, creating a ClearArray.
 852 // It writes zero bits in [start..end), within the body of an array object.
 853 // The memory effects are all chained onto the 'adr_type' alias category.
 854 //
 855 // Since the object is otherwise uninitialized, we are free
 856 // to put a little "slop" around the edges of the cleared area,
 857 // as long as it does not go back into the array's header,
 858 // or beyond the array end within the heap.
 859 //
 860 // The lower edge can be rounded down to the nearest jint and the
 861 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
 862 //
 863 // Arguments:
 864 //   adr_type           memory slice where writes are generated
 865 //   dest               oop of the destination array
 866 //   basic_elem_type    element type of the destination
 867 //   slice_idx          array index of first element to store
 868 //   slice_len          number of elements to store (or NULL)
 869 //   dest_size          total size in bytes of the array object
 870 //
 871 // Exactly one of slice_len or dest_size must be non-NULL.
 872 // If dest_size is non-NULL, zeroing extends to the end of the object.
 873 // If slice_len is non-NULL, the slice_idx value must be a constant.
 874 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
 875                                             const TypePtr* adr_type,
 876                                             Node* dest,


 877                                             BasicType basic_elem_type,
 878                                             Node* slice_idx,
 879                                             Node* slice_len,
 880                                             Node* dest_size) {
 881   // one or the other but not both of slice_len and dest_size:
 882   assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
 883   if (slice_len == NULL)  slice_len = top();
 884   if (dest_size == NULL)  dest_size = top();
 885 
 886   uint alias_idx = C->get_alias_index(adr_type);
 887 
 888   // operate on this memory slice:
 889   Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
 890 
 891   // scaling and rounding of indexes:

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






 963         mem = transform_later(mem);
 964       }
 965     }
 966     Node* end = dest_size; // pre-rounded
 967     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 968                                        start, end, &_igvn);
 969   } else {
 970     // Non-constant start, unrounded non-constant end.
 971     // (Nobody zeroes a random midsection of an array using this routine.)
 972     ShouldNotReachHere();       // fix caller
 973   }
 974 
 975   // Done.
 976   merge_mem->set_memory_at(alias_idx, mem);
 977 }
 978 
 979 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
 980                                                 const TypePtr* adr_type,
 981                                                 BasicType basic_elem_type,
 982                                                 AllocateNode* alloc,
 983                                                 Node* src,  Node* src_offset,
 984                                                 Node* dest, Node* dest_offset,
 985                                                 Node* dest_size, bool dest_uninitialized) {
 986   // See if there is an advantage from block transfer.
 987   int scale = exact_log2(type2aelembytes(basic_elem_type));
 988   if (scale >= LogBytesPerLong)
 989     return false;               // it is already a block transfer
 990 
 991   // Look at the alignment of the starting offsets.
 992   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 993 
 994   intptr_t src_off_con  = (intptr_t) _igvn.find_int_con(src_offset, -1);
 995   intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
 996   if (src_off_con < 0 || dest_off_con < 0) {
 997     // At present, we can only understand constants.
 998     return false;
 999   }
1000 
1001   intptr_t src_off  = abase + (src_off_con  << scale);
1002   intptr_t dest_off = abase + (dest_off_con << scale);
1003 
1004   if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
1005     // Non-aligned; too bad.
1006     // One more chance:  Pick off an initial 32-bit word.
1007     // This is a common case, since abase can be odd mod 8.
1008     if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
1009         ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
1010       Node* sptr = basic_plus_adr(src,  src_off);
1011       Node* dptr = basic_plus_adr(dest, dest_off);
1012       const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
1013       assert(s_adr_type->isa_aryptr(), "impossible slice");
1014       uint s_alias_idx = C->get_alias_index(s_adr_type);
1015       uint d_alias_idx = C->get_alias_index(adr_type);
1016       bool is_mismatched = (basic_elem_type != T_INT);
1017       Node* sval = transform_later(
1018           LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
1019                          TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
1020                          false /*unaligned*/, is_mismatched));
1021       Node* st = transform_later(
1022           StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
1023                           sval, T_INT, MemNode::unordered));
1024       if (is_mismatched) {
1025         st->as_Store()->set_mismatched_access();
1026       }
1027       (*mem)->set_memory_at(d_alias_idx, st);
1028       src_off += BytesPerInt;
1029       dest_off += BytesPerInt;
1030     } else {
1031       return false;
1032     }
1033   }
1034   assert(src_off % BytesPerLong == 0, "");
1035   assert(dest_off % BytesPerLong == 0, "");
1036 
1037   // Do this copy by giant steps.
1038   Node* sptr  = basic_plus_adr(src,  src_off);
1039   Node* dptr  = basic_plus_adr(dest, dest_off);
1040   Node* countx = dest_size;
1041   countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
1042   countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
1043 
1044   bool disjoint_bases = true;   // since alloc != NULL
1045   generate_unchecked_arraycopy(ctrl, mem,
1046                                adr_type, T_LONG, disjoint_bases,
1047                                sptr, NULL, dptr, NULL, countx, dest_uninitialized);
1048 
1049   return true;
1050 }
1051 
1052 // Helper function; generates code for the slow case.
1053 // We make a call to a runtime method which emulates the native method,
1054 // but without the native wrapper overhead.
1055 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
1056                                                         Node** ctrl, Node* mem, Node** io,
1057                                                         const TypePtr* adr_type,
1058                                                         Node* src,  Node* src_offset,
1059                                                         Node* dest, Node* dest_offset,
1060                                                         Node* copy_length, bool dest_uninitialized) {
1061   assert(!dest_uninitialized, "Invariant");
1062 
1063   const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
1064   CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
1065                                           "slow_arraycopy", TypePtr::BOTTOM);
1066 
1067   call->init_req(TypeFunc::Control, *ctrl);
1068   call->init_req(TypeFunc::I_O    , *io);
1069   call->init_req(TypeFunc::Memory , mem);
1070   call->init_req(TypeFunc::ReturnAdr, top());
1071   call->init_req(TypeFunc::FramePtr, top());
1072   call->init_req(TypeFunc::Parms+0, src);
1073   call->init_req(TypeFunc::Parms+1, src_offset);
1074   call->init_req(TypeFunc::Parms+2, dest);
1075   call->init_req(TypeFunc::Parms+3, dest_offset);
1076   call->init_req(TypeFunc::Parms+4, copy_length);
1077   call->copy_call_debug_info(&_igvn, ac);
1078 
1079   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
1080   _igvn.replace_node(ac, call);
1081   transform_later(call);
1082 
1083   call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1084   *ctrl = _callprojs.fallthrough_catchproj->clone();
1085   transform_later(*ctrl);
1086 
1087   Node* m = _callprojs.fallthrough_memproj->clone();
1088   transform_later(m);
1089 
1090   uint alias_idx = C->get_alias_index(adr_type);
1091   MergeMemNode* out_mem;
1092   if (alias_idx != Compile::AliasIdxBot) {
1093     out_mem = MergeMemNode::make(mem);
1094     out_mem->set_memory_at(alias_idx, m);
1095   } else {
1096     out_mem = MergeMemNode::make(m);
1097   }
1098   transform_later(out_mem);
1099 
1100   // When src is negative and arraycopy is before an infinite loop,_callprojs.fallthrough_ioproj
1101   // could be NULL. Skip clone and update NULL fallthrough_ioproj.
1102   if (_callprojs.fallthrough_ioproj != NULL) {
1103     *io = _callprojs.fallthrough_ioproj->clone();
1104     transform_later(*io);
1105   } else {
1106     *io = NULL;
1107   }
1108 
1109   return out_mem;
1110 }
1111 
1112 // Helper function; generates code for cases requiring runtime checks.
1113 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
1114                                                      const TypePtr* adr_type,
1115                                                      Node* dest_elem_klass,
1116                                                      Node* src,  Node* src_offset,
1117                                                      Node* dest, Node* dest_offset,
1118                                                      Node* copy_length, bool dest_uninitialized) {
1119   if ((*ctrl)->is_top())  return NULL;
1120 
1121   address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1122   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1123     return NULL;
1124   }
1125 
1126   // Pick out the parameters required to perform a store-check
1127   // for the target array.  This is an optimistic check.  It will
1128   // look in each non-null element's class, at the desired klass's
1129   // super_check_offset, for the desired klass.
1130   int sco_offset = in_bytes(Klass::super_check_offset_offset());
1131   Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1132   Node* n3 = new LoadINode(NULL, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1133   Node* check_offset = ConvI2X(transform_later(n3));
1134   Node* check_value  = dest_elem_klass;
1135 
1136   Node* src_start  = array_element_address(src,  src_offset,  T_OBJECT);
1137   Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1138 
1139   const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1140   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1141                               src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1142 
1143   finish_arraycopy_call(call, ctrl, mem, adr_type);
1144 
1145   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1146   transform_later(proj);
1147 
1148   return proj;
1149 }
1150 
1151 // Helper function; generates code for cases requiring runtime checks.
1152 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1153                                                    const TypePtr* adr_type,
1154                                                    Node* src,  Node* src_offset,
1155                                                    Node* dest, Node* dest_offset,
1156                                                    Node* copy_length, bool dest_uninitialized) {
1157   if ((*ctrl)->is_top()) return NULL;
1158   assert(!dest_uninitialized, "Invariant");
1159 
1160   address copyfunc_addr = StubRoutines::generic_arraycopy();
1161   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1162     return NULL;
1163   }
1164 
1165   const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1166   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1167                               src, src_offset, dest, dest_offset, copy_length);
1168 
1169   finish_arraycopy_call(call, ctrl, mem, adr_type);
1170 
1171   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1172   transform_later(proj);
1173 
1174   return proj;
1175 }
1176 
1177 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1178 bool PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1179                                                     const TypePtr* adr_type,
1180                                                     BasicType basic_elem_type,
1181                                                     bool disjoint_bases,
1182                                                     Node* src,  Node* src_offset,
1183                                                     Node* dest, Node* dest_offset,
1184                                                     Node* copy_length, bool dest_uninitialized) {
1185   if ((*ctrl)->is_top()) return false;
1186 
1187   Node* src_start  = src;
1188   Node* dest_start = dest;
1189   if (src_offset != NULL || dest_offset != NULL) {
1190     src_start =  array_element_address(src, src_offset, basic_elem_type);
1191     dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1192   }
1193 
1194   // Figure out which arraycopy runtime method to call.
1195   const char* copyfunc_name = "arraycopy";
1196   address     copyfunc_addr =
1197       basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1198                           disjoint_bases, copyfunc_name, dest_uninitialized);
1199 
1200   Node* result_memory = NULL;
1201   RegionNode* exit_block = NULL;
1202   if (ArrayOperationPartialInlineSize > 0 && is_subword_type(basic_elem_type) &&
1203     Matcher::vector_width_in_bytes(basic_elem_type) >= 16) {
1204     generate_partial_inlining_block(ctrl, mem, adr_type, &exit_block, &result_memory,
1205                                     copy_length, src_start, dest_start, basic_elem_type);
1206   }
1207 
1208   const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1209   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1210                               src_start, dest_start, copy_length XTOP);
1211 
1212   finish_arraycopy_call(call, ctrl, mem, adr_type);
1213 
1214   // Connecting remaining edges for exit_block coming from stub_block.
1215   if (exit_block) {
1216     exit_block->init_req(2, *ctrl);
1217 
1218     // Memory edge corresponding to stub_region.
1219     result_memory->init_req(2, *mem);
1220 
1221     uint alias_idx = C->get_alias_index(adr_type);
1222     if (alias_idx != Compile::AliasIdxBot) {
1223       *mem = MergeMemNode::make(*mem);
1224       (*mem)->set_memory_at(alias_idx, result_memory);
1225     } else {
1226       *mem = MergeMemNode::make(result_memory);
1227     }
1228     transform_later(*mem);
1229     *ctrl = exit_block;
1230     return true;
1231   }
1232   return false;
1233 }
1234 




































1235 #undef XTOP
1236 
1237 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1238   Node* ctrl = ac->in(TypeFunc::Control);
1239   Node* io = ac->in(TypeFunc::I_O);
1240   Node* src = ac->in(ArrayCopyNode::Src);
1241   Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1242   Node* dest = ac->in(ArrayCopyNode::Dest);
1243   Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1244   Node* length = ac->in(ArrayCopyNode::Length);
1245   MergeMemNode* merge_mem = NULL;
1246 
1247   if (ac->is_clonebasic()) {
1248     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1249     bs->clone_at_expansion(this, ac);
1250     return;
1251   } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_clone_oop_array()) {

















1252     Node* mem = ac->in(TypeFunc::Memory);
1253     merge_mem = MergeMemNode::make(mem);
1254     transform_later(merge_mem);
1255 
1256     AllocateArrayNode* alloc = NULL;

1257     if (ac->is_alloc_tightly_coupled()) {
1258       alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1259       assert(alloc != NULL, "expect alloc");

1260     }
1261 
1262     const TypePtr* adr_type = _igvn.type(dest)->is_oopptr()->add_offset(Type::OffsetBot);
1263     if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1264       adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();












1265     }
1266     generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1267                        adr_type, T_OBJECT,
1268                        src, src_offset, dest, dest_offset, length,

1269                        true, !ac->is_copyofrange());
1270 
1271     return;
1272   }
1273 
1274   AllocateArrayNode* alloc = NULL;
1275   if (ac->is_alloc_tightly_coupled()) {
1276     alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1277     assert(alloc != NULL, "expect alloc");
1278   }
1279 
1280   assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1281 
1282   // Compile time checks.  If any of these checks cannot be verified at compile time,
1283   // we do not make a fast path for this call.  Instead, we let the call remain as it
1284   // is.  The checks we choose to mandate at compile time are:
1285   //
1286   // (1) src and dest are arrays.
1287   const Type* src_type = src->Value(&_igvn);
1288   const Type* dest_type = dest->Value(&_igvn);
1289   const TypeAryPtr* top_src = src_type->isa_aryptr();
1290   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1291 
1292   BasicType src_elem = T_CONFLICT;
1293   BasicType dest_elem = T_CONFLICT;
1294 
1295   if (top_dest != NULL && top_dest->klass() != NULL) {
1296     dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
1297   }
1298   if (top_src != NULL && top_src->klass() != NULL) {
1299     src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
1300   }
1301   if (is_reference_type(src_elem))  src_elem  = T_OBJECT;
1302   if (is_reference_type(dest_elem)) dest_elem = T_OBJECT;




1303 
1304   if (ac->is_arraycopy_validated() &&
1305       dest_elem != T_CONFLICT &&
1306       src_elem == T_CONFLICT) {
1307     src_elem = dest_elem;
1308   }
1309 
1310   if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1311     // Conservatively insert a memory barrier on all memory slices.
1312     // Do not let writes into the source float below the arraycopy.
1313     {
1314       Node* mem = ac->in(TypeFunc::Memory);
1315       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1316 
1317       merge_mem = MergeMemNode::make(mem);
1318       transform_later(merge_mem);
1319     }
1320 
1321     // Call StubRoutines::generic_arraycopy stub.
1322     Node* mem = generate_arraycopy(ac, NULL, &ctrl, merge_mem, &io,
1323                                    TypeRawPtr::BOTTOM, T_CONFLICT,
1324                                    src, src_offset, dest, dest_offset, length,

1325                                    // If a  negative length guard was generated for the ArrayCopyNode,
1326                                    // the length of the array can never be negative.
1327                                    false, ac->has_negative_length_guard());
1328     return;
1329   }
1330 
1331   assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");

1332 
1333   // (2) src and dest arrays must have elements of the same BasicType
1334   // Figure out the size and type of the elements we will be copying.
1335   if (src_elem != dest_elem || dest_elem == T_VOID) {







1336     // The component types are not the same or are not recognized.  Punt.
1337     // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1338     {
1339       Node* mem = ac->in(TypeFunc::Memory);
1340       merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1341     }
1342 
1343     _igvn.replace_node(_callprojs.fallthrough_memproj, merge_mem);
1344     if (_callprojs.fallthrough_ioproj != NULL) {
1345       _igvn.replace_node(_callprojs.fallthrough_ioproj, io);
1346     }
1347     _igvn.replace_node(_callprojs.fallthrough_catchproj, ctrl);
1348     return;
1349   }
1350 
1351   //---------------------------------------------------------------------------
1352   // We will make a fast path for this call to arraycopy.
1353 
1354   // We have the following tests left to perform:
1355   //
1356   // (3) src and dest must not be null.
1357   // (4) src_offset must not be negative.
1358   // (5) dest_offset must not be negative.
1359   // (6) length must not be negative.
1360   // (7) src_offset + length must not exceed length of src.
1361   // (8) dest_offset + length must not exceed length of dest.
1362   // (9) each element of an oop array must be assignable
1363 
1364   {
1365     Node* mem = ac->in(TypeFunc::Memory);
1366     merge_mem = MergeMemNode::make(mem);
1367     transform_later(merge_mem);
1368   }
1369 
1370   RegionNode* slow_region = new RegionNode(1);
1371   transform_later(slow_region);
1372 
1373   if (!ac->is_arraycopy_validated()) {
1374     // (3) operands must not be null
1375     // We currently perform our null checks with the null_check routine.
1376     // This means that the null exceptions will be reported in the caller
1377     // rather than (correctly) reported inside of the native arraycopy call.
1378     // This should be corrected, given time.  We do our null check with the
1379     // stack pointer restored.
1380     // null checks done library_call.cpp
1381 
1382     // (4) src_offset must not be negative.
1383     generate_negative_guard(&ctrl, src_offset, slow_region);
1384 
1385     // (5) dest_offset must not be negative.
1386     generate_negative_guard(&ctrl, dest_offset, slow_region);
1387 
1388     // (6) length must not be negative (moved to generate_arraycopy()).
1389     // generate_negative_guard(length, slow_region);
1390 
1391     // (7) src_offset + length must not exceed length of src.
1392     Node* alen = ac->in(ArrayCopyNode::SrcLen);
1393     assert(alen != NULL, "need src len");
1394     generate_limit_guard(&ctrl,
1395                          src_offset, length,
1396                          alen,
1397                          slow_region);
1398 
1399     // (8) dest_offset + length must not exceed length of dest.
1400     alen = ac->in(ArrayCopyNode::DestLen);
1401     assert(alen != NULL, "need dest len");
1402     generate_limit_guard(&ctrl,
1403                          dest_offset, length,
1404                          alen,
1405                          slow_region);
1406 
1407     // (9) each element of an oop array must be assignable
1408     // The generate_arraycopy subroutine checks this.















1409   }

1410   // This is where the memory effects are placed:
1411   const TypePtr* adr_type = NULL;
1412   if (ac->_dest_type != TypeOopPtr::BOTTOM) {







1413     adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1414   } else {
1415     adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1416   }
1417 
1418   generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1419                      adr_type, dest_elem,
1420                      src, src_offset, dest, dest_offset, length,

1421                      // If a  negative length guard was generated for the ArrayCopyNode,
1422                      // the length of the array can never be negative.
1423                      false, ac->has_negative_length_guard(), slow_region);

1424 }
--- EOF ---