1 /* 2 * Copyright (c) 2007, 2012, 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 #include "precompiled.hpp" 25 #include "compiler/compileLog.hpp" 26 #include "libadt/vectset.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/divnode.hpp" 31 #include "opto/matcher.hpp" 32 #include "opto/memnode.hpp" 33 #include "opto/mulnode.hpp" 34 #include "opto/opcodes.hpp" 35 #include "opto/superword.hpp" 36 #include "opto/vectornode.hpp" 37 38 // 39 // S U P E R W O R D T R A N S F O R M 40 //============================================================================= 41 42 //------------------------------SuperWord--------------------------- 43 SuperWord::SuperWord(PhaseIdealLoop* phase) : 44 _phase(phase), 45 _igvn(phase->_igvn), 46 _arena(phase->C->comp_arena()), 47 _packset(arena(), 8, 0, NULL), // packs for the current block 48 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb 49 _block(arena(), 8, 0, NULL), // nodes in current block 50 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside 51 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads 52 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails 53 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node 54 _align_to_ref(NULL), // memory reference to align vectors to 55 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs 56 _dg(_arena), // dependence graph 57 _visited(arena()), // visited node set 58 _post_visited(arena()), // post visited node set 59 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs 60 _stk(arena(), 8, 0, NULL), // scratch stack of nodes 61 _nlist(arena(), 8, 0, NULL), // scratch list of nodes 62 _lpt(NULL), // loop tree node 63 _lp(NULL), // LoopNode 64 _bb(NULL), // basic block 65 _iv(NULL) // induction var 66 {} 67 68 //------------------------------transform_loop--------------------------- 69 void SuperWord::transform_loop(IdealLoopTree* lpt) { 70 assert(UseSuperWord, "should be"); 71 // Do vectors exist on this architecture? 72 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return; 73 74 assert(lpt->_head->is_CountedLoop(), "must be"); 75 CountedLoopNode *cl = lpt->_head->as_CountedLoop(); 76 77 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop 78 79 if (!cl->is_main_loop() ) return; // skip normal, pre, and post loops 80 81 // Check for no control flow in body (other than exit) 82 Node *cl_exit = cl->loopexit(); 83 if (cl_exit->in(0) != lpt->_head) return; 84 85 // Make sure the are no extra control users of the loop backedge 86 if (cl->back_control()->outcnt() != 1) { 87 return; 88 } 89 90 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit)))) 91 CountedLoopEndNode* pre_end = get_pre_loop_end(cl); 92 if (pre_end == NULL) return; 93 Node *pre_opaq1 = pre_end->limit(); 94 if (pre_opaq1->Opcode() != Op_Opaque1) return; 95 96 init(); // initialize data structures 97 98 set_lpt(lpt); 99 set_lp(cl); 100 101 // For now, define one block which is the entire loop body 102 set_bb(cl); 103 104 assert(_packset.length() == 0, "packset must be empty"); 105 SLP_extract(); 106 } 107 108 //------------------------------SLP_extract--------------------------- 109 // Extract the superword level parallelism 110 // 111 // 1) A reverse post-order of nodes in the block is constructed. By scanning 112 // this list from first to last, all definitions are visited before their uses. 113 // 114 // 2) A point-to-point dependence graph is constructed between memory references. 115 // This simplies the upcoming "independence" checker. 116 // 117 // 3) The maximum depth in the node graph from the beginning of the block 118 // to each node is computed. This is used to prune the graph search 119 // in the independence checker. 120 // 121 // 4) For integer types, the necessary bit width is propagated backwards 122 // from stores to allow packed operations on byte, char, and short 123 // integers. This reverses the promotion to type "int" that javac 124 // did for operations like: char c1,c2,c3; c1 = c2 + c3. 125 // 126 // 5) One of the memory references is picked to be an aligned vector reference. 127 // The pre-loop trip count is adjusted to align this reference in the 128 // unrolled body. 129 // 130 // 6) The initial set of pack pairs is seeded with memory references. 131 // 132 // 7) The set of pack pairs is extended by following use->def and def->use links. 133 // 134 // 8) The pairs are combined into vector sized packs. 135 // 136 // 9) Reorder the memory slices to co-locate members of the memory packs. 137 // 138 // 10) Generate ideal vector nodes for the final set of packs and where necessary, 139 // inserting scalar promotion, vector creation from multiple scalars, and 140 // extraction of scalar values from vectors. 141 // 142 void SuperWord::SLP_extract() { 143 144 // Ready the block 145 146 construct_bb(); 147 148 dependence_graph(); 149 150 compute_max_depth(); 151 152 compute_vector_element_type(); 153 154 // Attempt vectorization 155 156 find_adjacent_refs(); 157 158 extend_packlist(); 159 160 combine_packs(); 161 162 construct_my_pack_map(); 163 164 filter_packs(); 165 166 schedule(); 167 168 output(); 169 } 170 171 //------------------------------find_adjacent_refs--------------------------- 172 // Find the adjacent memory references and create pack pairs for them. 173 // This is the initial set of packs that will then be extended by 174 // following use->def and def->use links. The align positions are 175 // assigned relative to the reference "align_to_ref" 176 void SuperWord::find_adjacent_refs() { 177 // Get list of memory operations 178 Node_List memops; 179 for (int i = 0; i < _block.length(); i++) { 180 Node* n = _block.at(i); 181 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) && 182 n->Opcode() != Op_LoadUI2L && 183 is_java_primitive(n->as_Mem()->memory_type())) { 184 int align = memory_alignment(n->as_Mem(), 0); 185 if (align != bottom_align) { 186 memops.push(n); 187 } 188 } 189 } 190 191 Node_List align_to_refs; 192 int best_iv_adjustment = 0; 193 MemNode* best_align_to_mem_ref = NULL; 194 195 while (memops.size() != 0) { 196 // Find a memory reference to align to. 197 MemNode* mem_ref = find_align_to_ref(memops); 198 if (mem_ref == NULL) break; 199 align_to_refs.push(mem_ref); 200 int iv_adjustment = get_iv_adjustment(mem_ref); 201 202 if (best_align_to_mem_ref == NULL) { 203 // Set memory reference which is the best from all memory operations 204 // to be used for alignment. The pre-loop trip count is modified to align 205 // this reference to a vector-aligned address. 206 best_align_to_mem_ref = mem_ref; 207 best_iv_adjustment = iv_adjustment; 208 } 209 210 SWPointer align_to_ref_p(mem_ref, this); 211 // Set alignment relative to "align_to_ref" for all related memory operations. 212 for (int i = memops.size() - 1; i >= 0; i--) { 213 MemNode* s = memops.at(i)->as_Mem(); 214 if (isomorphic(s, mem_ref)) { 215 SWPointer p2(s, this); 216 if (p2.comparable(align_to_ref_p)) { 217 int align = memory_alignment(s, iv_adjustment); 218 set_alignment(s, align); 219 } 220 } 221 } 222 223 // Create initial pack pairs of memory operations for which 224 // alignment is set and vectors will be aligned. 225 bool create_pack = true; 226 if (memory_alignment(mem_ref, best_iv_adjustment) == 0) { 227 if (!Matcher::misaligned_vectors_ok()) { 228 int vw = vector_width(mem_ref); 229 int vw_best = vector_width(best_align_to_mem_ref); 230 if (vw > vw_best) { 231 // Do not vectorize a memory access with more elements per vector 232 // if unaligned memory access is not allowed because number of 233 // iterations in pre-loop will be not enough to align it. 234 create_pack = false; 235 } 236 } 237 } else { 238 if (same_velt_type(mem_ref, best_align_to_mem_ref)) { 239 // Can't allow vectorization of unaligned memory accesses with the 240 // same type since it could be overlapped accesses to the same array. 241 create_pack = false; 242 } else { 243 // Allow independent (different type) unaligned memory operations 244 // if HW supports them. 245 if (!Matcher::misaligned_vectors_ok()) { 246 create_pack = false; 247 } else { 248 // Check if packs of the same memory type but 249 // with a different alignment were created before. 250 for (uint i = 0; i < align_to_refs.size(); i++) { 251 MemNode* mr = align_to_refs.at(i)->as_Mem(); 252 if (same_velt_type(mr, mem_ref) && 253 memory_alignment(mr, iv_adjustment) != 0) 254 create_pack = false; 255 } 256 } 257 } 258 } 259 if (create_pack) { 260 for (uint i = 0; i < memops.size(); i++) { 261 Node* s1 = memops.at(i); 262 int align = alignment(s1); 263 if (align == top_align) continue; 264 for (uint j = 0; j < memops.size(); j++) { 265 Node* s2 = memops.at(j); 266 if (alignment(s2) == top_align) continue; 267 if (s1 != s2 && are_adjacent_refs(s1, s2)) { 268 if (stmts_can_pack(s1, s2, align)) { 269 Node_List* pair = new Node_List(); 270 pair->push(s1); 271 pair->push(s2); 272 _packset.append(pair); 273 } 274 } 275 } 276 } 277 } else { // Don't create unaligned pack 278 // First, remove remaining memory ops of the same type from the list. 279 for (int i = memops.size() - 1; i >= 0; i--) { 280 MemNode* s = memops.at(i)->as_Mem(); 281 if (same_velt_type(s, mem_ref)) { 282 memops.remove(i); 283 } 284 } 285 286 // Second, remove already constructed packs of the same type. 287 for (int i = _packset.length() - 1; i >= 0; i--) { 288 Node_List* p = _packset.at(i); 289 MemNode* s = p->at(0)->as_Mem(); 290 if (same_velt_type(s, mem_ref)) { 291 remove_pack_at(i); 292 } 293 } 294 295 // If needed find the best memory reference for loop alignment again. 296 if (same_velt_type(mem_ref, best_align_to_mem_ref)) { 297 // Put memory ops from remaining packs back on memops list for 298 // the best alignment search. 299 uint orig_msize = memops.size(); 300 for (int i = 0; i < _packset.length(); i++) { 301 Node_List* p = _packset.at(i); 302 MemNode* s = p->at(0)->as_Mem(); 303 assert(!same_velt_type(s, mem_ref), "sanity"); 304 memops.push(s); 305 } 306 MemNode* best_align_to_mem_ref = find_align_to_ref(memops); 307 if (best_align_to_mem_ref == NULL) break; 308 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref); 309 // Restore list. 310 while (memops.size() > orig_msize) 311 (void)memops.pop(); 312 } 313 } // unaligned memory accesses 314 315 // Remove used mem nodes. 316 for (int i = memops.size() - 1; i >= 0; i--) { 317 MemNode* m = memops.at(i)->as_Mem(); 318 if (alignment(m) != top_align) { 319 memops.remove(i); 320 } 321 } 322 323 } // while (memops.size() != 0 324 set_align_to_ref(best_align_to_mem_ref); 325 326 #ifndef PRODUCT 327 if (TraceSuperWord) { 328 tty->print_cr("\nAfter find_adjacent_refs"); 329 print_packset(); 330 } 331 #endif 332 } 333 334 //------------------------------find_align_to_ref--------------------------- 335 // Find a memory reference to align the loop induction variable to. 336 // Looks first at stores then at loads, looking for a memory reference 337 // with the largest number of references similar to it. 338 MemNode* SuperWord::find_align_to_ref(Node_List &memops) { 339 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0); 340 341 // Count number of comparable memory ops 342 for (uint i = 0; i < memops.size(); i++) { 343 MemNode* s1 = memops.at(i)->as_Mem(); 344 SWPointer p1(s1, this); 345 // Discard if pre loop can't align this reference 346 if (!ref_is_alignable(p1)) { 347 *cmp_ct.adr_at(i) = 0; 348 continue; 349 } 350 for (uint j = i+1; j < memops.size(); j++) { 351 MemNode* s2 = memops.at(j)->as_Mem(); 352 if (isomorphic(s1, s2)) { 353 SWPointer p2(s2, this); 354 if (p1.comparable(p2)) { 355 (*cmp_ct.adr_at(i))++; 356 (*cmp_ct.adr_at(j))++; 357 } 358 } 359 } 360 } 361 362 // Find Store (or Load) with the greatest number of "comparable" references, 363 // biggest vector size, smallest data size and smallest iv offset. 364 int max_ct = 0; 365 int max_vw = 0; 366 int max_idx = -1; 367 int min_size = max_jint; 368 int min_iv_offset = max_jint; 369 for (uint j = 0; j < memops.size(); j++) { 370 MemNode* s = memops.at(j)->as_Mem(); 371 if (s->is_Store()) { 372 int vw = vector_width_in_bytes(s); 373 assert(vw > 1, "sanity"); 374 SWPointer p(s, this); 375 if (cmp_ct.at(j) > max_ct || 376 cmp_ct.at(j) == max_ct && 377 (vw > max_vw || 378 vw == max_vw && 379 (data_size(s) < min_size || 380 data_size(s) == min_size && 381 (p.offset_in_bytes() < min_iv_offset)))) { 382 max_ct = cmp_ct.at(j); 383 max_vw = vw; 384 max_idx = j; 385 min_size = data_size(s); 386 min_iv_offset = p.offset_in_bytes(); 387 } 388 } 389 } 390 // If no stores, look at loads 391 if (max_ct == 0) { 392 for (uint j = 0; j < memops.size(); j++) { 393 MemNode* s = memops.at(j)->as_Mem(); 394 if (s->is_Load()) { 395 int vw = vector_width_in_bytes(s); 396 assert(vw > 1, "sanity"); 397 SWPointer p(s, this); 398 if (cmp_ct.at(j) > max_ct || 399 cmp_ct.at(j) == max_ct && 400 (vw > max_vw || 401 vw == max_vw && 402 (data_size(s) < min_size || 403 data_size(s) == min_size && 404 (p.offset_in_bytes() < min_iv_offset)))) { 405 max_ct = cmp_ct.at(j); 406 max_vw = vw; 407 max_idx = j; 408 min_size = data_size(s); 409 min_iv_offset = p.offset_in_bytes(); 410 } 411 } 412 } 413 } 414 415 #ifdef ASSERT 416 if (TraceSuperWord && Verbose) { 417 tty->print_cr("\nVector memops after find_align_to_refs"); 418 for (uint i = 0; i < memops.size(); i++) { 419 MemNode* s = memops.at(i)->as_Mem(); 420 s->dump(); 421 } 422 } 423 #endif 424 425 if (max_ct > 0) { 426 #ifdef ASSERT 427 if (TraceSuperWord) { 428 tty->print("\nVector align to node: "); 429 memops.at(max_idx)->as_Mem()->dump(); 430 } 431 #endif 432 return memops.at(max_idx)->as_Mem(); 433 } 434 return NULL; 435 } 436 437 //------------------------------ref_is_alignable--------------------------- 438 // Can the preloop align the reference to position zero in the vector? 439 bool SuperWord::ref_is_alignable(SWPointer& p) { 440 if (!p.has_iv()) { 441 return true; // no induction variable 442 } 443 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop()); 444 assert(pre_end->stride_is_con(), "pre loop stride is constant"); 445 int preloop_stride = pre_end->stride_con(); 446 447 int span = preloop_stride * p.scale_in_bytes(); 448 449 // Stride one accesses are alignable. 450 if (ABS(span) == p.memory_size()) 451 return true; 452 453 // If initial offset from start of object is computable, 454 // compute alignment within the vector. 455 int vw = vector_width_in_bytes(p.mem()); 456 assert(vw > 1, "sanity"); 457 if (vw % span == 0) { 458 Node* init_nd = pre_end->init_trip(); 459 if (init_nd->is_Con() && p.invar() == NULL) { 460 int init = init_nd->bottom_type()->is_int()->get_con(); 461 462 int init_offset = init * p.scale_in_bytes() + p.offset_in_bytes(); 463 assert(init_offset >= 0, "positive offset from object start"); 464 465 if (span > 0) { 466 return (vw - (init_offset % vw)) % span == 0; 467 } else { 468 assert(span < 0, "nonzero stride * scale"); 469 return (init_offset % vw) % -span == 0; 470 } 471 } 472 } 473 return false; 474 } 475 476 //---------------------------get_iv_adjustment--------------------------- 477 // Calculate loop's iv adjustment for this memory ops. 478 int SuperWord::get_iv_adjustment(MemNode* mem_ref) { 479 SWPointer align_to_ref_p(mem_ref, this); 480 int offset = align_to_ref_p.offset_in_bytes(); 481 int scale = align_to_ref_p.scale_in_bytes(); 482 int vw = vector_width_in_bytes(mem_ref); 483 assert(vw > 1, "sanity"); 484 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1; 485 // At least one iteration is executed in pre-loop by default. As result 486 // several iterations are needed to align memory operations in main-loop even 487 // if offset is 0. 488 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw)); 489 int elt_size = align_to_ref_p.memory_size(); 490 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0), 491 err_msg_res("(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size)); 492 int iv_adjustment = iv_adjustment_in_bytes/elt_size; 493 494 #ifndef PRODUCT 495 if (TraceSuperWord) 496 tty->print_cr("\noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d", 497 offset, iv_adjustment, elt_size, scale, iv_stride(), vw); 498 #endif 499 return iv_adjustment; 500 } 501 502 //---------------------------dependence_graph--------------------------- 503 // Construct dependency graph. 504 // Add dependence edges to load/store nodes for memory dependence 505 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x) 506 void SuperWord::dependence_graph() { 507 // First, assign a dependence node to each memory node 508 for (int i = 0; i < _block.length(); i++ ) { 509 Node *n = _block.at(i); 510 if (n->is_Mem() || n->is_Phi() && n->bottom_type() == Type::MEMORY) { 511 _dg.make_node(n); 512 } 513 } 514 515 // For each memory slice, create the dependences 516 for (int i = 0; i < _mem_slice_head.length(); i++) { 517 Node* n = _mem_slice_head.at(i); 518 Node* n_tail = _mem_slice_tail.at(i); 519 520 // Get slice in predecessor order (last is first) 521 mem_slice_preds(n_tail, n, _nlist); 522 523 // Make the slice dependent on the root 524 DepMem* slice = _dg.dep(n); 525 _dg.make_edge(_dg.root(), slice); 526 527 // Create a sink for the slice 528 DepMem* slice_sink = _dg.make_node(NULL); 529 _dg.make_edge(slice_sink, _dg.tail()); 530 531 // Now visit each pair of memory ops, creating the edges 532 for (int j = _nlist.length() - 1; j >= 0 ; j--) { 533 Node* s1 = _nlist.at(j); 534 535 // If no dependency yet, use slice 536 if (_dg.dep(s1)->in_cnt() == 0) { 537 _dg.make_edge(slice, s1); 538 } 539 SWPointer p1(s1->as_Mem(), this); 540 bool sink_dependent = true; 541 for (int k = j - 1; k >= 0; k--) { 542 Node* s2 = _nlist.at(k); 543 if (s1->is_Load() && s2->is_Load()) 544 continue; 545 SWPointer p2(s2->as_Mem(), this); 546 547 int cmp = p1.cmp(p2); 548 if (SuperWordRTDepCheck && 549 p1.base() != p2.base() && p1.valid() && p2.valid()) { 550 // Create a runtime check to disambiguate 551 OrderedPair pp(p1.base(), p2.base()); 552 _disjoint_ptrs.append_if_missing(pp); 553 } else if (!SWPointer::not_equal(cmp)) { 554 // Possibly same address 555 _dg.make_edge(s1, s2); 556 sink_dependent = false; 557 } 558 } 559 if (sink_dependent) { 560 _dg.make_edge(s1, slice_sink); 561 } 562 } 563 #ifndef PRODUCT 564 if (TraceSuperWord) { 565 tty->print_cr("\nDependence graph for slice: %d", n->_idx); 566 for (int q = 0; q < _nlist.length(); q++) { 567 _dg.print(_nlist.at(q)); 568 } 569 tty->cr(); 570 } 571 #endif 572 _nlist.clear(); 573 } 574 575 #ifndef PRODUCT 576 if (TraceSuperWord) { 577 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE"); 578 for (int r = 0; r < _disjoint_ptrs.length(); r++) { 579 _disjoint_ptrs.at(r).print(); 580 tty->cr(); 581 } 582 tty->cr(); 583 } 584 #endif 585 } 586 587 //---------------------------mem_slice_preds--------------------------- 588 // Return a memory slice (node list) in predecessor order starting at "start" 589 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) { 590 assert(preds.length() == 0, "start empty"); 591 Node* n = start; 592 Node* prev = NULL; 593 while (true) { 594 assert(in_bb(n), "must be in block"); 595 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 596 Node* out = n->fast_out(i); 597 if (out->is_Load()) { 598 if (in_bb(out)) { 599 preds.push(out); 600 } 601 } else { 602 // FIXME 603 if (out->is_MergeMem() && !in_bb(out)) { 604 // Either unrolling is causing a memory edge not to disappear, 605 // or need to run igvn.optimize() again before SLP 606 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) { 607 // Ditto. Not sure what else to check further. 608 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) { 609 // StoreCM has an input edge used as a precedence edge. 610 // Maybe an issue when oop stores are vectorized. 611 } else { 612 assert(out == prev || prev == NULL, "no branches off of store slice"); 613 } 614 } 615 } 616 if (n == stop) break; 617 preds.push(n); 618 prev = n; 619 n = n->in(MemNode::Memory); 620 } 621 } 622 623 //------------------------------stmts_can_pack--------------------------- 624 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and 625 // s1 aligned at "align" 626 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) { 627 628 // Do not use superword for non-primitives 629 BasicType bt1 = velt_basic_type(s1); 630 BasicType bt2 = velt_basic_type(s2); 631 if(!is_java_primitive(bt1) || !is_java_primitive(bt2)) 632 return false; 633 if (Matcher::max_vector_size(bt1) < 2) { 634 return false; // No vectors for this type 635 } 636 637 if (isomorphic(s1, s2)) { 638 if (independent(s1, s2)) { 639 if (!exists_at(s1, 0) && !exists_at(s2, 1)) { 640 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) { 641 int s1_align = alignment(s1); 642 int s2_align = alignment(s2); 643 if (s1_align == top_align || s1_align == align) { 644 if (s2_align == top_align || s2_align == align + data_size(s1)) { 645 return true; 646 } 647 } 648 } 649 } 650 } 651 } 652 return false; 653 } 654 655 //------------------------------exists_at--------------------------- 656 // Does s exist in a pack at position pos? 657 bool SuperWord::exists_at(Node* s, uint pos) { 658 for (int i = 0; i < _packset.length(); i++) { 659 Node_List* p = _packset.at(i); 660 if (p->at(pos) == s) { 661 return true; 662 } 663 } 664 return false; 665 } 666 667 //------------------------------are_adjacent_refs--------------------------- 668 // Is s1 immediately before s2 in memory? 669 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) { 670 if (!s1->is_Mem() || !s2->is_Mem()) return false; 671 if (!in_bb(s1) || !in_bb(s2)) return false; 672 673 // Do not use superword for non-primitives 674 if (!is_java_primitive(s1->as_Mem()->memory_type()) || 675 !is_java_primitive(s2->as_Mem()->memory_type())) { 676 return false; 677 } 678 679 // FIXME - co_locate_pack fails on Stores in different mem-slices, so 680 // only pack memops that are in the same alias set until that's fixed. 681 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) != 682 _phase->C->get_alias_index(s2->as_Mem()->adr_type())) 683 return false; 684 SWPointer p1(s1->as_Mem(), this); 685 SWPointer p2(s2->as_Mem(), this); 686 if (p1.base() != p2.base() || !p1.comparable(p2)) return false; 687 int diff = p2.offset_in_bytes() - p1.offset_in_bytes(); 688 return diff == data_size(s1); 689 } 690 691 //------------------------------isomorphic--------------------------- 692 // Are s1 and s2 similar? 693 bool SuperWord::isomorphic(Node* s1, Node* s2) { 694 if (s1->Opcode() != s2->Opcode()) return false; 695 if (s1->req() != s2->req()) return false; 696 if (s1->in(0) != s2->in(0)) return false; 697 if (!same_velt_type(s1, s2)) return false; 698 return true; 699 } 700 701 //------------------------------independent--------------------------- 702 // Is there no data path from s1 to s2 or s2 to s1? 703 bool SuperWord::independent(Node* s1, Node* s2) { 704 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first"); 705 int d1 = depth(s1); 706 int d2 = depth(s2); 707 if (d1 == d2) return s1 != s2; 708 Node* deep = d1 > d2 ? s1 : s2; 709 Node* shallow = d1 > d2 ? s2 : s1; 710 711 visited_clear(); 712 713 return independent_path(shallow, deep); 714 } 715 716 //------------------------------independent_path------------------------------ 717 // Helper for independent 718 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) { 719 if (dp >= 1000) return false; // stop deep recursion 720 visited_set(deep); 721 int shal_depth = depth(shallow); 722 assert(shal_depth <= depth(deep), "must be"); 723 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) { 724 Node* pred = preds.current(); 725 if (in_bb(pred) && !visited_test(pred)) { 726 if (shallow == pred) { 727 return false; 728 } 729 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) { 730 return false; 731 } 732 } 733 } 734 return true; 735 } 736 737 //------------------------------set_alignment--------------------------- 738 void SuperWord::set_alignment(Node* s1, Node* s2, int align) { 739 set_alignment(s1, align); 740 if (align == top_align || align == bottom_align) { 741 set_alignment(s2, align); 742 } else { 743 set_alignment(s2, align + data_size(s1)); 744 } 745 } 746 747 //------------------------------data_size--------------------------- 748 int SuperWord::data_size(Node* s) { 749 int bsize = type2aelembytes(velt_basic_type(s)); 750 assert(bsize != 0, "valid size"); 751 return bsize; 752 } 753 754 //------------------------------extend_packlist--------------------------- 755 // Extend packset by following use->def and def->use links from pack members. 756 void SuperWord::extend_packlist() { 757 bool changed; 758 do { 759 changed = false; 760 for (int i = 0; i < _packset.length(); i++) { 761 Node_List* p = _packset.at(i); 762 changed |= follow_use_defs(p); 763 changed |= follow_def_uses(p); 764 } 765 } while (changed); 766 767 #ifndef PRODUCT 768 if (TraceSuperWord) { 769 tty->print_cr("\nAfter extend_packlist"); 770 print_packset(); 771 } 772 #endif 773 } 774 775 //------------------------------follow_use_defs--------------------------- 776 // Extend the packset by visiting operand definitions of nodes in pack p 777 bool SuperWord::follow_use_defs(Node_List* p) { 778 assert(p->size() == 2, "just checking"); 779 Node* s1 = p->at(0); 780 Node* s2 = p->at(1); 781 assert(s1->req() == s2->req(), "just checking"); 782 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking"); 783 784 if (s1->is_Load()) return false; 785 786 int align = alignment(s1); 787 bool changed = false; 788 int start = s1->is_Store() ? MemNode::ValueIn : 1; 789 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req(); 790 for (int j = start; j < end; j++) { 791 Node* t1 = s1->in(j); 792 Node* t2 = s2->in(j); 793 if (!in_bb(t1) || !in_bb(t2)) 794 continue; 795 if (stmts_can_pack(t1, t2, align)) { 796 if (est_savings(t1, t2) >= 0) { 797 Node_List* pair = new Node_List(); 798 pair->push(t1); 799 pair->push(t2); 800 _packset.append(pair); 801 set_alignment(t1, t2, align); 802 changed = true; 803 } 804 } 805 } 806 return changed; 807 } 808 809 //------------------------------follow_def_uses--------------------------- 810 // Extend the packset by visiting uses of nodes in pack p 811 bool SuperWord::follow_def_uses(Node_List* p) { 812 bool changed = false; 813 Node* s1 = p->at(0); 814 Node* s2 = p->at(1); 815 assert(p->size() == 2, "just checking"); 816 assert(s1->req() == s2->req(), "just checking"); 817 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking"); 818 819 if (s1->is_Store()) return false; 820 821 int align = alignment(s1); 822 int savings = -1; 823 Node* u1 = NULL; 824 Node* u2 = NULL; 825 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { 826 Node* t1 = s1->fast_out(i); 827 if (!in_bb(t1)) continue; 828 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) { 829 Node* t2 = s2->fast_out(j); 830 if (!in_bb(t2)) continue; 831 if (!opnd_positions_match(s1, t1, s2, t2)) 832 continue; 833 if (stmts_can_pack(t1, t2, align)) { 834 int my_savings = est_savings(t1, t2); 835 if (my_savings > savings) { 836 savings = my_savings; 837 u1 = t1; 838 u2 = t2; 839 } 840 } 841 } 842 } 843 if (savings >= 0) { 844 Node_List* pair = new Node_List(); 845 pair->push(u1); 846 pair->push(u2); 847 _packset.append(pair); 848 set_alignment(u1, u2, align); 849 changed = true; 850 } 851 return changed; 852 } 853 854 //---------------------------opnd_positions_match------------------------- 855 // Is the use of d1 in u1 at the same operand position as d2 in u2? 856 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) { 857 uint ct = u1->req(); 858 if (ct != u2->req()) return false; 859 uint i1 = 0; 860 uint i2 = 0; 861 do { 862 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break; 863 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break; 864 if (i1 != i2) { 865 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) { 866 // Further analysis relies on operands position matching. 867 u2->swap_edges(i1, i2); 868 } else { 869 return false; 870 } 871 } 872 } while (i1 < ct); 873 return true; 874 } 875 876 //------------------------------est_savings--------------------------- 877 // Estimate the savings from executing s1 and s2 as a pack 878 int SuperWord::est_savings(Node* s1, Node* s2) { 879 int save_in = 2 - 1; // 2 operations per instruction in packed form 880 881 // inputs 882 for (uint i = 1; i < s1->req(); i++) { 883 Node* x1 = s1->in(i); 884 Node* x2 = s2->in(i); 885 if (x1 != x2) { 886 if (are_adjacent_refs(x1, x2)) { 887 save_in += adjacent_profit(x1, x2); 888 } else if (!in_packset(x1, x2)) { 889 save_in -= pack_cost(2); 890 } else { 891 save_in += unpack_cost(2); 892 } 893 } 894 } 895 896 // uses of result 897 uint ct = 0; 898 int save_use = 0; 899 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { 900 Node* s1_use = s1->fast_out(i); 901 for (int j = 0; j < _packset.length(); j++) { 902 Node_List* p = _packset.at(j); 903 if (p->at(0) == s1_use) { 904 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) { 905 Node* s2_use = s2->fast_out(k); 906 if (p->at(p->size()-1) == s2_use) { 907 ct++; 908 if (are_adjacent_refs(s1_use, s2_use)) { 909 save_use += adjacent_profit(s1_use, s2_use); 910 } 911 } 912 } 913 } 914 } 915 } 916 917 if (ct < s1->outcnt()) save_use += unpack_cost(1); 918 if (ct < s2->outcnt()) save_use += unpack_cost(1); 919 920 return MAX2(save_in, save_use); 921 } 922 923 //------------------------------costs--------------------------- 924 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; } 925 int SuperWord::pack_cost(int ct) { return ct; } 926 int SuperWord::unpack_cost(int ct) { return ct; } 927 928 //------------------------------combine_packs--------------------------- 929 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last 930 void SuperWord::combine_packs() { 931 bool changed = true; 932 // Combine packs regardless max vector size. 933 while (changed) { 934 changed = false; 935 for (int i = 0; i < _packset.length(); i++) { 936 Node_List* p1 = _packset.at(i); 937 if (p1 == NULL) continue; 938 for (int j = 0; j < _packset.length(); j++) { 939 Node_List* p2 = _packset.at(j); 940 if (p2 == NULL) continue; 941 if (i == j) continue; 942 if (p1->at(p1->size()-1) == p2->at(0)) { 943 for (uint k = 1; k < p2->size(); k++) { 944 p1->push(p2->at(k)); 945 } 946 _packset.at_put(j, NULL); 947 changed = true; 948 } 949 } 950 } 951 } 952 953 // Split packs which have size greater then max vector size. 954 for (int i = 0; i < _packset.length(); i++) { 955 Node_List* p1 = _packset.at(i); 956 if (p1 != NULL) { 957 BasicType bt = velt_basic_type(p1->at(0)); 958 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector 959 assert(is_power_of_2(max_vlen), "sanity"); 960 uint psize = p1->size(); 961 if (!is_power_of_2(psize)) { 962 // Skip pack which can't be vector. 963 // case1: for(...) { a[i] = i; } elements values are different (i+x) 964 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store 965 _packset.at_put(i, NULL); 966 continue; 967 } 968 if (psize > max_vlen) { 969 Node_List* pack = new Node_List(); 970 for (uint j = 0; j < psize; j++) { 971 pack->push(p1->at(j)); 972 if (pack->size() >= max_vlen) { 973 assert(is_power_of_2(pack->size()), "sanity"); 974 _packset.append(pack); 975 pack = new Node_List(); 976 } 977 } 978 _packset.at_put(i, NULL); 979 } 980 } 981 } 982 983 // Compress list. 984 for (int i = _packset.length() - 1; i >= 0; i--) { 985 Node_List* p1 = _packset.at(i); 986 if (p1 == NULL) { 987 _packset.remove_at(i); 988 } 989 } 990 991 #ifndef PRODUCT 992 if (TraceSuperWord) { 993 tty->print_cr("\nAfter combine_packs"); 994 print_packset(); 995 } 996 #endif 997 } 998 999 //-----------------------------construct_my_pack_map-------------------------- 1000 // Construct the map from nodes to packs. Only valid after the 1001 // point where a node is only in one pack (after combine_packs). 1002 void SuperWord::construct_my_pack_map() { 1003 Node_List* rslt = NULL; 1004 for (int i = 0; i < _packset.length(); i++) { 1005 Node_List* p = _packset.at(i); 1006 for (uint j = 0; j < p->size(); j++) { 1007 Node* s = p->at(j); 1008 assert(my_pack(s) == NULL, "only in one pack"); 1009 set_my_pack(s, p); 1010 } 1011 } 1012 } 1013 1014 //------------------------------filter_packs--------------------------- 1015 // Remove packs that are not implemented or not profitable. 1016 void SuperWord::filter_packs() { 1017 1018 // Remove packs that are not implemented 1019 for (int i = _packset.length() - 1; i >= 0; i--) { 1020 Node_List* pk = _packset.at(i); 1021 bool impl = implemented(pk); 1022 if (!impl) { 1023 #ifndef PRODUCT 1024 if (TraceSuperWord && Verbose) { 1025 tty->print_cr("Unimplemented"); 1026 pk->at(0)->dump(); 1027 } 1028 #endif 1029 remove_pack_at(i); 1030 } 1031 } 1032 1033 // Remove packs that are not profitable 1034 bool changed; 1035 do { 1036 changed = false; 1037 for (int i = _packset.length() - 1; i >= 0; i--) { 1038 Node_List* pk = _packset.at(i); 1039 bool prof = profitable(pk); 1040 if (!prof) { 1041 #ifndef PRODUCT 1042 if (TraceSuperWord && Verbose) { 1043 tty->print_cr("Unprofitable"); 1044 pk->at(0)->dump(); 1045 } 1046 #endif 1047 remove_pack_at(i); 1048 changed = true; 1049 } 1050 } 1051 } while (changed); 1052 1053 #ifndef PRODUCT 1054 if (TraceSuperWord) { 1055 tty->print_cr("\nAfter filter_packs"); 1056 print_packset(); 1057 tty->cr(); 1058 } 1059 #endif 1060 } 1061 1062 //------------------------------implemented--------------------------- 1063 // Can code be generated for pack p? 1064 bool SuperWord::implemented(Node_List* p) { 1065 Node* p0 = p->at(0); 1066 return VectorNode::implemented(p0->Opcode(), p->size(), velt_basic_type(p0)); 1067 } 1068 1069 //------------------------------same_inputs-------------------------- 1070 // For pack p, are all idx operands the same? 1071 static bool same_inputs(Node_List* p, int idx) { 1072 Node* p0 = p->at(0); 1073 uint vlen = p->size(); 1074 Node* p0_def = p0->in(idx); 1075 for (uint i = 1; i < vlen; i++) { 1076 Node* pi = p->at(i); 1077 Node* pi_def = pi->in(idx); 1078 if (p0_def != pi_def) 1079 return false; 1080 } 1081 return true; 1082 } 1083 1084 //------------------------------profitable--------------------------- 1085 // For pack p, are all operands and all uses (with in the block) vector? 1086 bool SuperWord::profitable(Node_List* p) { 1087 Node* p0 = p->at(0); 1088 uint start, end; 1089 VectorNode::vector_operands(p0, &start, &end); 1090 1091 // Return false if some inputs are not vectors or vectors with different 1092 // size or alignment. 1093 // Also, for now, return false if not scalar promotion case when inputs are 1094 // the same. Later, implement PackNode and allow differing, non-vector inputs 1095 // (maybe just the ones from outside the block.) 1096 for (uint i = start; i < end; i++) { 1097 if (!is_vector_use(p0, i)) 1098 return false; 1099 } 1100 if (VectorNode::is_shift(p0)) { 1101 // For now, return false if shift count is vector or not scalar promotion 1102 // case (different shift counts) because it is not supported yet. 1103 Node* cnt = p0->in(2); 1104 Node_List* cnt_pk = my_pack(cnt); 1105 if (cnt_pk != NULL) 1106 return false; 1107 if (!same_inputs(p, 2)) 1108 return false; 1109 } 1110 if (!p0->is_Store()) { 1111 // For now, return false if not all uses are vector. 1112 // Later, implement ExtractNode and allow non-vector uses (maybe 1113 // just the ones outside the block.) 1114 for (uint i = 0; i < p->size(); i++) { 1115 Node* def = p->at(i); 1116 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 1117 Node* use = def->fast_out(j); 1118 for (uint k = 0; k < use->req(); k++) { 1119 Node* n = use->in(k); 1120 if (def == n) { 1121 if (!is_vector_use(use, k)) { 1122 return false; 1123 } 1124 } 1125 } 1126 } 1127 } 1128 } 1129 return true; 1130 } 1131 1132 //------------------------------schedule--------------------------- 1133 // Adjust the memory graph for the packed operations 1134 void SuperWord::schedule() { 1135 1136 // Co-locate in the memory graph the members of each memory pack 1137 for (int i = 0; i < _packset.length(); i++) { 1138 co_locate_pack(_packset.at(i)); 1139 } 1140 } 1141 1142 //-------------------------------remove_and_insert------------------- 1143 // Remove "current" from its current position in the memory graph and insert 1144 // it after the appropriate insertion point (lip or uip). 1145 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip, 1146 Node *uip, Unique_Node_List &sched_before) { 1147 Node* my_mem = current->in(MemNode::Memory); 1148 bool sched_up = sched_before.member(current); 1149 1150 // remove current_store from its current position in the memmory graph 1151 for (DUIterator i = current->outs(); current->has_out(i); i++) { 1152 Node* use = current->out(i); 1153 if (use->is_Mem()) { 1154 assert(use->in(MemNode::Memory) == current, "must be"); 1155 if (use == prev) { // connect prev to my_mem 1156 _igvn.replace_input_of(use, MemNode::Memory, my_mem); 1157 --i; //deleted this edge; rescan position 1158 } else if (sched_before.member(use)) { 1159 if (!sched_up) { // Will be moved together with current 1160 _igvn.replace_input_of(use, MemNode::Memory, uip); 1161 --i; //deleted this edge; rescan position 1162 } 1163 } else { 1164 if (sched_up) { // Will be moved together with current 1165 _igvn.replace_input_of(use, MemNode::Memory, lip); 1166 --i; //deleted this edge; rescan position 1167 } 1168 } 1169 } 1170 } 1171 1172 Node *insert_pt = sched_up ? uip : lip; 1173 1174 // all uses of insert_pt's memory state should use current's instead 1175 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) { 1176 Node* use = insert_pt->out(i); 1177 if (use->is_Mem()) { 1178 assert(use->in(MemNode::Memory) == insert_pt, "must be"); 1179 _igvn.replace_input_of(use, MemNode::Memory, current); 1180 --i; //deleted this edge; rescan position 1181 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) { 1182 uint pos; //lip (lower insert point) must be the last one in the memory slice 1183 for (pos=1; pos < use->req(); pos++) { 1184 if (use->in(pos) == insert_pt) break; 1185 } 1186 _igvn.replace_input_of(use, pos, current); 1187 --i; 1188 } 1189 } 1190 1191 //connect current to insert_pt 1192 _igvn.replace_input_of(current, MemNode::Memory, insert_pt); 1193 } 1194 1195 //------------------------------co_locate_pack---------------------------------- 1196 // To schedule a store pack, we need to move any sandwiched memory ops either before 1197 // or after the pack, based upon dependence information: 1198 // (1) If any store in the pack depends on the sandwiched memory op, the 1199 // sandwiched memory op must be scheduled BEFORE the pack; 1200 // (2) If a sandwiched memory op depends on any store in the pack, the 1201 // sandwiched memory op must be scheduled AFTER the pack; 1202 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched 1203 // memory op (say memB), memB must be scheduled before memA. So, if memA is 1204 // scheduled before the pack, memB must also be scheduled before the pack; 1205 // (4) If there is no dependence restriction for a sandwiched memory op, we simply 1206 // schedule this store AFTER the pack 1207 // (5) We know there is no dependence cycle, so there in no other case; 1208 // (6) Finally, all memory ops in another single pack should be moved in the same direction. 1209 // 1210 // To schedule a load pack, we use the memory state of either the first or the last load in 1211 // the pack, based on the dependence constraint. 1212 void SuperWord::co_locate_pack(Node_List* pk) { 1213 if (pk->at(0)->is_Store()) { 1214 MemNode* first = executed_first(pk)->as_Mem(); 1215 MemNode* last = executed_last(pk)->as_Mem(); 1216 Unique_Node_List schedule_before_pack; 1217 Unique_Node_List memops; 1218 1219 MemNode* current = last->in(MemNode::Memory)->as_Mem(); 1220 MemNode* previous = last; 1221 while (true) { 1222 assert(in_bb(current), "stay in block"); 1223 memops.push(previous); 1224 for (DUIterator i = current->outs(); current->has_out(i); i++) { 1225 Node* use = current->out(i); 1226 if (use->is_Mem() && use != previous) 1227 memops.push(use); 1228 } 1229 if (current == first) break; 1230 previous = current; 1231 current = current->in(MemNode::Memory)->as_Mem(); 1232 } 1233 1234 // determine which memory operations should be scheduled before the pack 1235 for (uint i = 1; i < memops.size(); i++) { 1236 Node *s1 = memops.at(i); 1237 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) { 1238 for (uint j = 0; j< i; j++) { 1239 Node *s2 = memops.at(j); 1240 if (!independent(s1, s2)) { 1241 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) { 1242 schedule_before_pack.push(s1); // s1 must be scheduled before 1243 Node_List* mem_pk = my_pack(s1); 1244 if (mem_pk != NULL) { 1245 for (uint ii = 0; ii < mem_pk->size(); ii++) { 1246 Node* s = mem_pk->at(ii); // follow partner 1247 if (memops.member(s) && !schedule_before_pack.member(s)) 1248 schedule_before_pack.push(s); 1249 } 1250 } 1251 break; 1252 } 1253 } 1254 } 1255 } 1256 } 1257 1258 Node* upper_insert_pt = first->in(MemNode::Memory); 1259 // Following code moves loads connected to upper_insert_pt below aliased stores. 1260 // Collect such loads here and reconnect them back to upper_insert_pt later. 1261 memops.clear(); 1262 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) { 1263 Node* use = upper_insert_pt->out(i); 1264 if (!use->is_Store()) 1265 memops.push(use); 1266 } 1267 1268 MemNode* lower_insert_pt = last; 1269 previous = last; //previous store in pk 1270 current = last->in(MemNode::Memory)->as_Mem(); 1271 1272 // start scheduling from "last" to "first" 1273 while (true) { 1274 assert(in_bb(current), "stay in block"); 1275 assert(in_pack(previous, pk), "previous stays in pack"); 1276 Node* my_mem = current->in(MemNode::Memory); 1277 1278 if (in_pack(current, pk)) { 1279 // Forward users of my memory state (except "previous) to my input memory state 1280 for (DUIterator i = current->outs(); current->has_out(i); i++) { 1281 Node* use = current->out(i); 1282 if (use->is_Mem() && use != previous) { 1283 assert(use->in(MemNode::Memory) == current, "must be"); 1284 if (schedule_before_pack.member(use)) { 1285 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt); 1286 } else { 1287 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt); 1288 } 1289 --i; // deleted this edge; rescan position 1290 } 1291 } 1292 previous = current; 1293 } else { // !in_pack(current, pk) ==> a sandwiched store 1294 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack); 1295 } 1296 1297 if (current == first) break; 1298 current = my_mem->as_Mem(); 1299 } // end while 1300 1301 // Reconnect loads back to upper_insert_pt. 1302 for (uint i = 0; i < memops.size(); i++) { 1303 Node *ld = memops.at(i); 1304 if (ld->in(MemNode::Memory) != upper_insert_pt) { 1305 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt); 1306 } 1307 } 1308 } else if (pk->at(0)->is_Load()) { //load 1309 // all loads in the pack should have the same memory state. By default, 1310 // we use the memory state of the last load. However, if any load could 1311 // not be moved down due to the dependence constraint, we use the memory 1312 // state of the first load. 1313 Node* last_mem = executed_last(pk)->in(MemNode::Memory); 1314 Node* first_mem = executed_first(pk)->in(MemNode::Memory); 1315 bool schedule_last = true; 1316 for (uint i = 0; i < pk->size(); i++) { 1317 Node* ld = pk->at(i); 1318 for (Node* current = last_mem; current != ld->in(MemNode::Memory); 1319 current=current->in(MemNode::Memory)) { 1320 assert(current != first_mem, "corrupted memory graph"); 1321 if(current->is_Mem() && !independent(current, ld)){ 1322 schedule_last = false; // a later store depends on this load 1323 break; 1324 } 1325 } 1326 } 1327 1328 Node* mem_input = schedule_last ? last_mem : first_mem; 1329 _igvn.hash_delete(mem_input); 1330 // Give each load the same memory state 1331 for (uint i = 0; i < pk->size(); i++) { 1332 LoadNode* ld = pk->at(i)->as_Load(); 1333 _igvn.replace_input_of(ld, MemNode::Memory, mem_input); 1334 } 1335 } 1336 } 1337 1338 //------------------------------output--------------------------- 1339 // Convert packs into vector node operations 1340 void SuperWord::output() { 1341 if (_packset.length() == 0) return; 1342 1343 #ifndef PRODUCT 1344 if (TraceLoopOpts) { 1345 tty->print("SuperWord "); 1346 lpt()->dump_head(); 1347 } 1348 #endif 1349 1350 // MUST ENSURE main loop's initial value is properly aligned: 1351 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0 1352 1353 align_initial_loop_index(align_to_ref()); 1354 1355 // Insert extract (unpack) operations for scalar uses 1356 for (int i = 0; i < _packset.length(); i++) { 1357 insert_extracts(_packset.at(i)); 1358 } 1359 1360 Compile* C = _phase->C; 1361 uint max_vlen_in_bytes = 0; 1362 for (int i = 0; i < _block.length(); i++) { 1363 Node* n = _block.at(i); 1364 Node_List* p = my_pack(n); 1365 if (p && n == executed_last(p)) { 1366 uint vlen = p->size(); 1367 uint vlen_in_bytes = 0; 1368 Node* vn = NULL; 1369 Node* low_adr = p->at(0); 1370 Node* first = executed_first(p); 1371 int opc = n->Opcode(); 1372 if (n->is_Load()) { 1373 Node* ctl = n->in(MemNode::Control); 1374 Node* mem = first->in(MemNode::Memory); 1375 Node* adr = low_adr->in(MemNode::Address); 1376 const TypePtr* atyp = n->adr_type(); 1377 vn = LoadVectorNode::make(C, opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n)); 1378 vlen_in_bytes = vn->as_LoadVector()->memory_size(); 1379 } else if (n->is_Store()) { 1380 // Promote value to be stored to vector 1381 Node* val = vector_opd(p, MemNode::ValueIn); 1382 Node* ctl = n->in(MemNode::Control); 1383 Node* mem = first->in(MemNode::Memory); 1384 Node* adr = low_adr->in(MemNode::Address); 1385 const TypePtr* atyp = n->adr_type(); 1386 vn = StoreVectorNode::make(C, opc, ctl, mem, adr, atyp, val, vlen); 1387 vlen_in_bytes = vn->as_StoreVector()->memory_size(); 1388 } else if (n->req() == 3) { 1389 // Promote operands to vector 1390 Node* in1 = vector_opd(p, 1); 1391 Node* in2 = vector_opd(p, 2); 1392 if (VectorNode::is_invariant_vector(in1) && (n->is_Add() || n->is_Mul())) { 1393 // Move invariant vector input into second position to avoid register spilling. 1394 Node* tmp = in1; 1395 in1 = in2; 1396 in2 = tmp; 1397 } 1398 vn = VectorNode::make(C, opc, in1, in2, vlen, velt_basic_type(n)); 1399 vlen_in_bytes = vn->as_Vector()->length_in_bytes(); 1400 } else { 1401 ShouldNotReachHere(); 1402 } 1403 assert(vn != NULL, "sanity"); 1404 _igvn.register_new_node_with_optimizer(vn); 1405 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0))); 1406 for (uint j = 0; j < p->size(); j++) { 1407 Node* pm = p->at(j); 1408 _igvn.replace_node(pm, vn); 1409 } 1410 _igvn._worklist.push(vn); 1411 1412 if (vlen_in_bytes > max_vlen_in_bytes) { 1413 max_vlen_in_bytes = vlen_in_bytes; 1414 } 1415 #ifdef ASSERT 1416 if (TraceNewVectors) { 1417 tty->print("new Vector node: "); 1418 vn->dump(); 1419 } 1420 #endif 1421 } 1422 } 1423 C->set_max_vector_size(max_vlen_in_bytes); 1424 } 1425 1426 //------------------------------vector_opd--------------------------- 1427 // Create a vector operand for the nodes in pack p for operand: in(opd_idx) 1428 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) { 1429 Node* p0 = p->at(0); 1430 uint vlen = p->size(); 1431 Node* opd = p0->in(opd_idx); 1432 1433 if (same_inputs(p, opd_idx)) { 1434 if (opd->is_Vector() || opd->is_LoadVector()) { 1435 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector"); 1436 return opd; // input is matching vector 1437 } 1438 if ((opd_idx == 2) && VectorNode::is_shift(p0)) { 1439 Compile* C = _phase->C; 1440 Node* cnt = opd; 1441 // Vector instructions do not mask shift count, do it here. 1442 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 1443 const TypeInt* t = opd->find_int_type(); 1444 if (t != NULL && t->is_con()) { 1445 juint shift = t->get_con(); 1446 if (shift > mask) { // Unsigned cmp 1447 cnt = ConNode::make(C, TypeInt::make(shift & mask)); 1448 } 1449 } else { 1450 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 1451 cnt = ConNode::make(C, TypeInt::make(mask)); 1452 _igvn.register_new_node_with_optimizer(cnt); 1453 cnt = new (C, 3) AndINode(opd, cnt); 1454 _igvn.register_new_node_with_optimizer(cnt); 1455 _phase->set_ctrl(cnt, _phase->get_ctrl(opd)); 1456 } 1457 assert(opd->bottom_type()->isa_int(), "int type only"); 1458 // Move non constant shift count into vector register. 1459 cnt = VectorNode::shift_count(C, p0, cnt, vlen, velt_basic_type(p0)); 1460 } 1461 if (cnt != opd) { 1462 _igvn.register_new_node_with_optimizer(cnt); 1463 _phase->set_ctrl(cnt, _phase->get_ctrl(opd)); 1464 } 1465 return cnt; 1466 } 1467 assert(!opd->is_StoreVector(), "such vector is not expected here"); 1468 // Convert scalar input to vector with the same number of elements as 1469 // p0's vector. Use p0's type because size of operand's container in 1470 // vector should match p0's size regardless operand's size. 1471 const Type* p0_t = velt_type(p0); 1472 VectorNode* vn = VectorNode::scalar2vector(_phase->C, opd, vlen, p0_t); 1473 1474 _igvn.register_new_node_with_optimizer(vn); 1475 _phase->set_ctrl(vn, _phase->get_ctrl(opd)); 1476 #ifdef ASSERT 1477 if (TraceNewVectors) { 1478 tty->print("new Vector node: "); 1479 vn->dump(); 1480 } 1481 #endif 1482 return vn; 1483 } 1484 1485 // Insert pack operation 1486 BasicType bt = velt_basic_type(p0); 1487 PackNode* pk = PackNode::make(_phase->C, opd, vlen, bt); 1488 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); ) 1489 1490 for (uint i = 1; i < vlen; i++) { 1491 Node* pi = p->at(i); 1492 Node* in = pi->in(opd_idx); 1493 assert(my_pack(in) == NULL, "Should already have been unpacked"); 1494 assert(opd_bt == in->bottom_type()->basic_type(), "all same type"); 1495 pk->add_opd(in); 1496 } 1497 _igvn.register_new_node_with_optimizer(pk); 1498 _phase->set_ctrl(pk, _phase->get_ctrl(opd)); 1499 #ifdef ASSERT 1500 if (TraceNewVectors) { 1501 tty->print("new Vector node: "); 1502 pk->dump(); 1503 } 1504 #endif 1505 return pk; 1506 } 1507 1508 //------------------------------insert_extracts--------------------------- 1509 // If a use of pack p is not a vector use, then replace the 1510 // use with an extract operation. 1511 void SuperWord::insert_extracts(Node_List* p) { 1512 if (p->at(0)->is_Store()) return; 1513 assert(_n_idx_list.is_empty(), "empty (node,index) list"); 1514 1515 // Inspect each use of each pack member. For each use that is 1516 // not a vector use, replace the use with an extract operation. 1517 1518 for (uint i = 0; i < p->size(); i++) { 1519 Node* def = p->at(i); 1520 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 1521 Node* use = def->fast_out(j); 1522 for (uint k = 0; k < use->req(); k++) { 1523 Node* n = use->in(k); 1524 if (def == n) { 1525 if (!is_vector_use(use, k)) { 1526 _n_idx_list.push(use, k); 1527 } 1528 } 1529 } 1530 } 1531 } 1532 1533 while (_n_idx_list.is_nonempty()) { 1534 Node* use = _n_idx_list.node(); 1535 int idx = _n_idx_list.index(); 1536 _n_idx_list.pop(); 1537 Node* def = use->in(idx); 1538 1539 // Insert extract operation 1540 _igvn.hash_delete(def); 1541 int def_pos = alignment(def) / data_size(def); 1542 1543 Node* ex = ExtractNode::make(_phase->C, def, def_pos, velt_basic_type(def)); 1544 _igvn.register_new_node_with_optimizer(ex); 1545 _phase->set_ctrl(ex, _phase->get_ctrl(def)); 1546 _igvn.replace_input_of(use, idx, ex); 1547 _igvn._worklist.push(def); 1548 1549 bb_insert_after(ex, bb_idx(def)); 1550 set_velt_type(ex, velt_type(def)); 1551 } 1552 } 1553 1554 //------------------------------is_vector_use--------------------------- 1555 // Is use->in(u_idx) a vector use? 1556 bool SuperWord::is_vector_use(Node* use, int u_idx) { 1557 Node_List* u_pk = my_pack(use); 1558 if (u_pk == NULL) return false; 1559 Node* def = use->in(u_idx); 1560 Node_List* d_pk = my_pack(def); 1561 if (d_pk == NULL) { 1562 // check for scalar promotion 1563 Node* n = u_pk->at(0)->in(u_idx); 1564 for (uint i = 1; i < u_pk->size(); i++) { 1565 if (u_pk->at(i)->in(u_idx) != n) return false; 1566 } 1567 return true; 1568 } 1569 if (u_pk->size() != d_pk->size()) 1570 return false; 1571 for (uint i = 0; i < u_pk->size(); i++) { 1572 Node* ui = u_pk->at(i); 1573 Node* di = d_pk->at(i); 1574 if (ui->in(u_idx) != di || alignment(ui) != alignment(di)) 1575 return false; 1576 } 1577 return true; 1578 } 1579 1580 //------------------------------construct_bb--------------------------- 1581 // Construct reverse postorder list of block members 1582 void SuperWord::construct_bb() { 1583 Node* entry = bb(); 1584 1585 assert(_stk.length() == 0, "stk is empty"); 1586 assert(_block.length() == 0, "block is empty"); 1587 assert(_data_entry.length() == 0, "data_entry is empty"); 1588 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty"); 1589 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty"); 1590 1591 // Find non-control nodes with no inputs from within block, 1592 // create a temporary map from node _idx to bb_idx for use 1593 // by the visited and post_visited sets, 1594 // and count number of nodes in block. 1595 int bb_ct = 0; 1596 for (uint i = 0; i < lpt()->_body.size(); i++ ) { 1597 Node *n = lpt()->_body.at(i); 1598 set_bb_idx(n, i); // Create a temporary map 1599 if (in_bb(n)) { 1600 bb_ct++; 1601 if (!n->is_CFG()) { 1602 bool found = false; 1603 for (uint j = 0; j < n->req(); j++) { 1604 Node* def = n->in(j); 1605 if (def && in_bb(def)) { 1606 found = true; 1607 break; 1608 } 1609 } 1610 if (!found) { 1611 assert(n != entry, "can't be entry"); 1612 _data_entry.push(n); 1613 } 1614 } 1615 } 1616 } 1617 1618 // Find memory slices (head and tail) 1619 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) { 1620 Node *n = lp()->fast_out(i); 1621 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) { 1622 Node* n_tail = n->in(LoopNode::LoopBackControl); 1623 if (n_tail != n->in(LoopNode::EntryControl)) { 1624 _mem_slice_head.push(n); 1625 _mem_slice_tail.push(n_tail); 1626 } 1627 } 1628 } 1629 1630 // Create an RPO list of nodes in block 1631 1632 visited_clear(); 1633 post_visited_clear(); 1634 1635 // Push all non-control nodes with no inputs from within block, then control entry 1636 for (int j = 0; j < _data_entry.length(); j++) { 1637 Node* n = _data_entry.at(j); 1638 visited_set(n); 1639 _stk.push(n); 1640 } 1641 visited_set(entry); 1642 _stk.push(entry); 1643 1644 // Do a depth first walk over out edges 1645 int rpo_idx = bb_ct - 1; 1646 int size; 1647 while ((size = _stk.length()) > 0) { 1648 Node* n = _stk.top(); // Leave node on stack 1649 if (!visited_test_set(n)) { 1650 // forward arc in graph 1651 } else if (!post_visited_test(n)) { 1652 // cross or back arc 1653 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1654 Node *use = n->fast_out(i); 1655 if (in_bb(use) && !visited_test(use) && 1656 // Don't go around backedge 1657 (!use->is_Phi() || n == entry)) { 1658 _stk.push(use); 1659 } 1660 } 1661 if (_stk.length() == size) { 1662 // There were no additional uses, post visit node now 1663 _stk.pop(); // Remove node from stack 1664 assert(rpo_idx >= 0, ""); 1665 _block.at_put_grow(rpo_idx, n); 1666 rpo_idx--; 1667 post_visited_set(n); 1668 assert(rpo_idx >= 0 || _stk.is_empty(), ""); 1669 } 1670 } else { 1671 _stk.pop(); // Remove post-visited node from stack 1672 } 1673 } 1674 1675 // Create real map of block indices for nodes 1676 for (int j = 0; j < _block.length(); j++) { 1677 Node* n = _block.at(j); 1678 set_bb_idx(n, j); 1679 } 1680 1681 initialize_bb(); // Ensure extra info is allocated. 1682 1683 #ifndef PRODUCT 1684 if (TraceSuperWord) { 1685 print_bb(); 1686 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE"); 1687 for (int m = 0; m < _data_entry.length(); m++) { 1688 tty->print("%3d ", m); 1689 _data_entry.at(m)->dump(); 1690 } 1691 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE"); 1692 for (int m = 0; m < _mem_slice_head.length(); m++) { 1693 tty->print("%3d ", m); _mem_slice_head.at(m)->dump(); 1694 tty->print(" "); _mem_slice_tail.at(m)->dump(); 1695 } 1696 } 1697 #endif 1698 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found"); 1699 } 1700 1701 //------------------------------initialize_bb--------------------------- 1702 // Initialize per node info 1703 void SuperWord::initialize_bb() { 1704 Node* last = _block.at(_block.length() - 1); 1705 grow_node_info(bb_idx(last)); 1706 } 1707 1708 //------------------------------bb_insert_after--------------------------- 1709 // Insert n into block after pos 1710 void SuperWord::bb_insert_after(Node* n, int pos) { 1711 int n_pos = pos + 1; 1712 // Make room 1713 for (int i = _block.length() - 1; i >= n_pos; i--) { 1714 _block.at_put_grow(i+1, _block.at(i)); 1715 } 1716 for (int j = _node_info.length() - 1; j >= n_pos; j--) { 1717 _node_info.at_put_grow(j+1, _node_info.at(j)); 1718 } 1719 // Set value 1720 _block.at_put_grow(n_pos, n); 1721 _node_info.at_put_grow(n_pos, SWNodeInfo::initial); 1722 // Adjust map from node->_idx to _block index 1723 for (int i = n_pos; i < _block.length(); i++) { 1724 set_bb_idx(_block.at(i), i); 1725 } 1726 } 1727 1728 //------------------------------compute_max_depth--------------------------- 1729 // Compute max depth for expressions from beginning of block 1730 // Use to prune search paths during test for independence. 1731 void SuperWord::compute_max_depth() { 1732 int ct = 0; 1733 bool again; 1734 do { 1735 again = false; 1736 for (int i = 0; i < _block.length(); i++) { 1737 Node* n = _block.at(i); 1738 if (!n->is_Phi()) { 1739 int d_orig = depth(n); 1740 int d_in = 0; 1741 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) { 1742 Node* pred = preds.current(); 1743 if (in_bb(pred)) { 1744 d_in = MAX2(d_in, depth(pred)); 1745 } 1746 } 1747 if (d_in + 1 != d_orig) { 1748 set_depth(n, d_in + 1); 1749 again = true; 1750 } 1751 } 1752 } 1753 ct++; 1754 } while (again); 1755 #ifndef PRODUCT 1756 if (TraceSuperWord && Verbose) 1757 tty->print_cr("compute_max_depth iterated: %d times", ct); 1758 #endif 1759 } 1760 1761 //-------------------------compute_vector_element_type----------------------- 1762 // Compute necessary vector element type for expressions 1763 // This propagates backwards a narrower integer type when the 1764 // upper bits of the value are not needed. 1765 // Example: char a,b,c; a = b + c; 1766 // Normally the type of the add is integer, but for packed character 1767 // operations the type of the add needs to be char. 1768 void SuperWord::compute_vector_element_type() { 1769 #ifndef PRODUCT 1770 if (TraceSuperWord && Verbose) 1771 tty->print_cr("\ncompute_velt_type:"); 1772 #endif 1773 1774 // Initial type 1775 for (int i = 0; i < _block.length(); i++) { 1776 Node* n = _block.at(i); 1777 set_velt_type(n, container_type(n)); 1778 } 1779 1780 // Propagate narrowed type backwards through operations 1781 // that don't depend on higher order bits 1782 for (int i = _block.length() - 1; i >= 0; i--) { 1783 Node* n = _block.at(i); 1784 // Only integer types need be examined 1785 const Type* vt = velt_type(n); 1786 if (vt->basic_type() == T_INT) { 1787 uint start, end; 1788 VectorNode::vector_operands(n, &start, &end); 1789 const Type* vt = velt_type(n); 1790 1791 for (uint j = start; j < end; j++) { 1792 Node* in = n->in(j); 1793 // Don't propagate through a memory 1794 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT && 1795 data_size(n) < data_size(in)) { 1796 bool same_type = true; 1797 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) { 1798 Node *use = in->fast_out(k); 1799 if (!in_bb(use) || !same_velt_type(use, n)) { 1800 same_type = false; 1801 break; 1802 } 1803 } 1804 if (same_type) { 1805 set_velt_type(in, vt); 1806 } 1807 } 1808 } 1809 } 1810 } 1811 #ifndef PRODUCT 1812 if (TraceSuperWord && Verbose) { 1813 for (int i = 0; i < _block.length(); i++) { 1814 Node* n = _block.at(i); 1815 velt_type(n)->dump(); 1816 tty->print("\t"); 1817 n->dump(); 1818 } 1819 } 1820 #endif 1821 } 1822 1823 //------------------------------memory_alignment--------------------------- 1824 // Alignment within a vector memory reference 1825 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) { 1826 SWPointer p(s, this); 1827 if (!p.valid()) { 1828 return bottom_align; 1829 } 1830 int vw = vector_width_in_bytes(s); 1831 if (vw < 2) { 1832 return bottom_align; // No vectors for this type 1833 } 1834 int offset = p.offset_in_bytes(); 1835 offset += iv_adjust*p.memory_size(); 1836 int off_rem = offset % vw; 1837 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw; 1838 return off_mod; 1839 } 1840 1841 //---------------------------container_type--------------------------- 1842 // Smallest type containing range of values 1843 const Type* SuperWord::container_type(Node* n) { 1844 if (n->is_Mem()) { 1845 return Type::get_const_basic_type(n->as_Mem()->memory_type()); 1846 } 1847 const Type* t = _igvn.type(n); 1848 if (t->basic_type() == T_INT) { 1849 // A narrow type of arithmetic operations will be determined by 1850 // propagating the type of memory operations. 1851 return TypeInt::INT; 1852 } 1853 return t; 1854 } 1855 1856 bool SuperWord::same_velt_type(Node* n1, Node* n2) { 1857 const Type* vt1 = velt_type(n1); 1858 const Type* vt2 = velt_type(n2); 1859 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) { 1860 // Compare vectors element sizes for integer types. 1861 return data_size(n1) == data_size(n2); 1862 } 1863 return vt1 == vt2; 1864 } 1865 1866 //------------------------------in_packset--------------------------- 1867 // Are s1 and s2 in a pack pair and ordered as s1,s2? 1868 bool SuperWord::in_packset(Node* s1, Node* s2) { 1869 for (int i = 0; i < _packset.length(); i++) { 1870 Node_List* p = _packset.at(i); 1871 assert(p->size() == 2, "must be"); 1872 if (p->at(0) == s1 && p->at(p->size()-1) == s2) { 1873 return true; 1874 } 1875 } 1876 return false; 1877 } 1878 1879 //------------------------------in_pack--------------------------- 1880 // Is s in pack p? 1881 Node_List* SuperWord::in_pack(Node* s, Node_List* p) { 1882 for (uint i = 0; i < p->size(); i++) { 1883 if (p->at(i) == s) { 1884 return p; 1885 } 1886 } 1887 return NULL; 1888 } 1889 1890 //------------------------------remove_pack_at--------------------------- 1891 // Remove the pack at position pos in the packset 1892 void SuperWord::remove_pack_at(int pos) { 1893 Node_List* p = _packset.at(pos); 1894 for (uint i = 0; i < p->size(); i++) { 1895 Node* s = p->at(i); 1896 set_my_pack(s, NULL); 1897 } 1898 _packset.remove_at(pos); 1899 } 1900 1901 //------------------------------executed_first--------------------------- 1902 // Return the node executed first in pack p. Uses the RPO block list 1903 // to determine order. 1904 Node* SuperWord::executed_first(Node_List* p) { 1905 Node* n = p->at(0); 1906 int n_rpo = bb_idx(n); 1907 for (uint i = 1; i < p->size(); i++) { 1908 Node* s = p->at(i); 1909 int s_rpo = bb_idx(s); 1910 if (s_rpo < n_rpo) { 1911 n = s; 1912 n_rpo = s_rpo; 1913 } 1914 } 1915 return n; 1916 } 1917 1918 //------------------------------executed_last--------------------------- 1919 // Return the node executed last in pack p. 1920 Node* SuperWord::executed_last(Node_List* p) { 1921 Node* n = p->at(0); 1922 int n_rpo = bb_idx(n); 1923 for (uint i = 1; i < p->size(); i++) { 1924 Node* s = p->at(i); 1925 int s_rpo = bb_idx(s); 1926 if (s_rpo > n_rpo) { 1927 n = s; 1928 n_rpo = s_rpo; 1929 } 1930 } 1931 return n; 1932 } 1933 1934 //----------------------------align_initial_loop_index--------------------------- 1935 // Adjust pre-loop limit so that in main loop, a load/store reference 1936 // to align_to_ref will be a position zero in the vector. 1937 // (iv + k) mod vector_align == 0 1938 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) { 1939 CountedLoopNode *main_head = lp()->as_CountedLoop(); 1940 assert(main_head->is_main_loop(), ""); 1941 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head); 1942 assert(pre_end != NULL, ""); 1943 Node *pre_opaq1 = pre_end->limit(); 1944 assert(pre_opaq1->Opcode() == Op_Opaque1, ""); 1945 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 1946 Node *lim0 = pre_opaq->in(1); 1947 1948 // Where we put new limit calculations 1949 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 1950 1951 // Ensure the original loop limit is available from the 1952 // pre-loop Opaque1 node. 1953 Node *orig_limit = pre_opaq->original_loop_limit(); 1954 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, ""); 1955 1956 SWPointer align_to_ref_p(align_to_ref, this); 1957 assert(align_to_ref_p.valid(), "sanity"); 1958 1959 // Given: 1960 // lim0 == original pre loop limit 1961 // V == v_align (power of 2) 1962 // invar == extra invariant piece of the address expression 1963 // e == offset [ +/- invar ] 1964 // 1965 // When reassociating expressions involving '%' the basic rules are: 1966 // (a - b) % k == 0 => a % k == b % k 1967 // and: 1968 // (a + b) % k == 0 => a % k == (k - b) % k 1969 // 1970 // For stride > 0 && scale > 0, 1971 // Derive the new pre-loop limit "lim" such that the two constraints: 1972 // (1) lim = lim0 + N (where N is some positive integer < V) 1973 // (2) (e + lim) % V == 0 1974 // are true. 1975 // 1976 // Substituting (1) into (2), 1977 // (e + lim0 + N) % V == 0 1978 // solve for N: 1979 // N = (V - (e + lim0)) % V 1980 // substitute back into (1), so that new limit 1981 // lim = lim0 + (V - (e + lim0)) % V 1982 // 1983 // For stride > 0 && scale < 0 1984 // Constraints: 1985 // lim = lim0 + N 1986 // (e - lim) % V == 0 1987 // Solving for lim: 1988 // (e - lim0 - N) % V == 0 1989 // N = (e - lim0) % V 1990 // lim = lim0 + (e - lim0) % V 1991 // 1992 // For stride < 0 && scale > 0 1993 // Constraints: 1994 // lim = lim0 - N 1995 // (e + lim) % V == 0 1996 // Solving for lim: 1997 // (e + lim0 - N) % V == 0 1998 // N = (e + lim0) % V 1999 // lim = lim0 - (e + lim0) % V 2000 // 2001 // For stride < 0 && scale < 0 2002 // Constraints: 2003 // lim = lim0 - N 2004 // (e - lim) % V == 0 2005 // Solving for lim: 2006 // (e - lim0 + N) % V == 0 2007 // N = (V - (e - lim0)) % V 2008 // lim = lim0 - (V - (e - lim0)) % V 2009 2010 int vw = vector_width_in_bytes(align_to_ref); 2011 int stride = iv_stride(); 2012 int scale = align_to_ref_p.scale_in_bytes(); 2013 int elt_size = align_to_ref_p.memory_size(); 2014 int v_align = vw / elt_size; 2015 assert(v_align > 1, "sanity"); 2016 int offset = align_to_ref_p.offset_in_bytes() / elt_size; 2017 Node *offsn = _igvn.intcon(offset); 2018 2019 Node *e = offsn; 2020 if (align_to_ref_p.invar() != NULL) { 2021 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt) 2022 Node* log2_elt = _igvn.intcon(exact_log2(elt_size)); 2023 Node* aref = new (_phase->C, 3) URShiftINode(align_to_ref_p.invar(), log2_elt); 2024 _igvn.register_new_node_with_optimizer(aref); 2025 _phase->set_ctrl(aref, pre_ctrl); 2026 if (align_to_ref_p.negate_invar()) { 2027 e = new (_phase->C, 3) SubINode(e, aref); 2028 } else { 2029 e = new (_phase->C, 3) AddINode(e, aref); 2030 } 2031 _igvn.register_new_node_with_optimizer(e); 2032 _phase->set_ctrl(e, pre_ctrl); 2033 } 2034 if (vw > ObjectAlignmentInBytes) { 2035 // incorporate base e +/- base && Mask >>> log2(elt) 2036 Node* xbase = new(_phase->C, 2) CastP2XNode(NULL, align_to_ref_p.base()); 2037 _igvn.register_new_node_with_optimizer(xbase); 2038 #ifdef _LP64 2039 xbase = new (_phase->C, 2) ConvL2INode(xbase); 2040 _igvn.register_new_node_with_optimizer(xbase); 2041 #endif 2042 Node* mask = _igvn.intcon(vw-1); 2043 Node* masked_xbase = new (_phase->C, 3) AndINode(xbase, mask); 2044 _igvn.register_new_node_with_optimizer(masked_xbase); 2045 Node* log2_elt = _igvn.intcon(exact_log2(elt_size)); 2046 Node* bref = new (_phase->C, 3) URShiftINode(masked_xbase, log2_elt); 2047 _igvn.register_new_node_with_optimizer(bref); 2048 _phase->set_ctrl(bref, pre_ctrl); 2049 e = new (_phase->C, 3) AddINode(e, bref); 2050 _igvn.register_new_node_with_optimizer(e); 2051 _phase->set_ctrl(e, pre_ctrl); 2052 } 2053 2054 // compute e +/- lim0 2055 if (scale < 0) { 2056 e = new (_phase->C, 3) SubINode(e, lim0); 2057 } else { 2058 e = new (_phase->C, 3) AddINode(e, lim0); 2059 } 2060 _igvn.register_new_node_with_optimizer(e); 2061 _phase->set_ctrl(e, pre_ctrl); 2062 2063 if (stride * scale > 0) { 2064 // compute V - (e +/- lim0) 2065 Node* va = _igvn.intcon(v_align); 2066 e = new (_phase->C, 3) SubINode(va, e); 2067 _igvn.register_new_node_with_optimizer(e); 2068 _phase->set_ctrl(e, pre_ctrl); 2069 } 2070 // compute N = (exp) % V 2071 Node* va_msk = _igvn.intcon(v_align - 1); 2072 Node* N = new (_phase->C, 3) AndINode(e, va_msk); 2073 _igvn.register_new_node_with_optimizer(N); 2074 _phase->set_ctrl(N, pre_ctrl); 2075 2076 // substitute back into (1), so that new limit 2077 // lim = lim0 + N 2078 Node* lim; 2079 if (stride < 0) { 2080 lim = new (_phase->C, 3) SubINode(lim0, N); 2081 } else { 2082 lim = new (_phase->C, 3) AddINode(lim0, N); 2083 } 2084 _igvn.register_new_node_with_optimizer(lim); 2085 _phase->set_ctrl(lim, pre_ctrl); 2086 Node* constrained = 2087 (stride > 0) ? (Node*) new (_phase->C,3) MinINode(lim, orig_limit) 2088 : (Node*) new (_phase->C,3) MaxINode(lim, orig_limit); 2089 _igvn.register_new_node_with_optimizer(constrained); 2090 _phase->set_ctrl(constrained, pre_ctrl); 2091 _igvn.hash_delete(pre_opaq); 2092 pre_opaq->set_req(1, constrained); 2093 } 2094 2095 //----------------------------get_pre_loop_end--------------------------- 2096 // Find pre loop end from main loop. Returns null if none. 2097 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode *cl) { 2098 Node *ctrl = cl->in(LoopNode::EntryControl); 2099 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return NULL; 2100 Node *iffm = ctrl->in(0); 2101 if (!iffm->is_If()) return NULL; 2102 Node *p_f = iffm->in(0); 2103 if (!p_f->is_IfFalse()) return NULL; 2104 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL; 2105 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2106 if (!pre_end->loopnode()->is_pre_loop()) return NULL; 2107 return pre_end; 2108 } 2109 2110 2111 //------------------------------init--------------------------- 2112 void SuperWord::init() { 2113 _dg.init(); 2114 _packset.clear(); 2115 _disjoint_ptrs.clear(); 2116 _block.clear(); 2117 _data_entry.clear(); 2118 _mem_slice_head.clear(); 2119 _mem_slice_tail.clear(); 2120 _node_info.clear(); 2121 _align_to_ref = NULL; 2122 _lpt = NULL; 2123 _lp = NULL; 2124 _bb = NULL; 2125 _iv = NULL; 2126 } 2127 2128 //------------------------------print_packset--------------------------- 2129 void SuperWord::print_packset() { 2130 #ifndef PRODUCT 2131 tty->print_cr("packset"); 2132 for (int i = 0; i < _packset.length(); i++) { 2133 tty->print_cr("Pack: %d", i); 2134 Node_List* p = _packset.at(i); 2135 print_pack(p); 2136 } 2137 #endif 2138 } 2139 2140 //------------------------------print_pack--------------------------- 2141 void SuperWord::print_pack(Node_List* p) { 2142 for (uint i = 0; i < p->size(); i++) { 2143 print_stmt(p->at(i)); 2144 } 2145 } 2146 2147 //------------------------------print_bb--------------------------- 2148 void SuperWord::print_bb() { 2149 #ifndef PRODUCT 2150 tty->print_cr("\nBlock"); 2151 for (int i = 0; i < _block.length(); i++) { 2152 Node* n = _block.at(i); 2153 tty->print("%d ", i); 2154 if (n) { 2155 n->dump(); 2156 } 2157 } 2158 #endif 2159 } 2160 2161 //------------------------------print_stmt--------------------------- 2162 void SuperWord::print_stmt(Node* s) { 2163 #ifndef PRODUCT 2164 tty->print(" align: %d \t", alignment(s)); 2165 s->dump(); 2166 #endif 2167 } 2168 2169 //------------------------------blank--------------------------- 2170 char* SuperWord::blank(uint depth) { 2171 static char blanks[101]; 2172 assert(depth < 101, "too deep"); 2173 for (uint i = 0; i < depth; i++) blanks[i] = ' '; 2174 blanks[depth] = '\0'; 2175 return blanks; 2176 } 2177 2178 2179 //==============================SWPointer=========================== 2180 2181 //----------------------------SWPointer------------------------ 2182 SWPointer::SWPointer(MemNode* mem, SuperWord* slp) : 2183 _mem(mem), _slp(slp), _base(NULL), _adr(NULL), 2184 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) { 2185 2186 Node* adr = mem->in(MemNode::Address); 2187 if (!adr->is_AddP()) { 2188 assert(!valid(), "too complex"); 2189 return; 2190 } 2191 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant) 2192 Node* base = adr->in(AddPNode::Base); 2193 //unsafe reference could not be aligned appropriately without runtime checking 2194 if (base == NULL || base->bottom_type() == Type::TOP) { 2195 assert(!valid(), "unsafe access"); 2196 return; 2197 } 2198 for (int i = 0; i < 3; i++) { 2199 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) { 2200 assert(!valid(), "too complex"); 2201 return; 2202 } 2203 adr = adr->in(AddPNode::Address); 2204 if (base == adr || !adr->is_AddP()) { 2205 break; // stop looking at addp's 2206 } 2207 } 2208 _base = base; 2209 _adr = adr; 2210 assert(valid(), "Usable"); 2211 } 2212 2213 // Following is used to create a temporary object during 2214 // the pattern match of an address expression. 2215 SWPointer::SWPointer(SWPointer* p) : 2216 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL), 2217 _scale(0), _offset(0), _invar(NULL), _negate_invar(false) {} 2218 2219 //------------------------scaled_iv_plus_offset-------------------- 2220 // Match: k*iv + offset 2221 // where: k is a constant that maybe zero, and 2222 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional 2223 bool SWPointer::scaled_iv_plus_offset(Node* n) { 2224 if (scaled_iv(n)) { 2225 return true; 2226 } 2227 if (offset_plus_k(n)) { 2228 return true; 2229 } 2230 int opc = n->Opcode(); 2231 if (opc == Op_AddI) { 2232 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) { 2233 return true; 2234 } 2235 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) { 2236 return true; 2237 } 2238 } else if (opc == Op_SubI) { 2239 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) { 2240 return true; 2241 } 2242 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) { 2243 _scale *= -1; 2244 return true; 2245 } 2246 } 2247 return false; 2248 } 2249 2250 //----------------------------scaled_iv------------------------ 2251 // Match: k*iv where k is a constant that's not zero 2252 bool SWPointer::scaled_iv(Node* n) { 2253 if (_scale != 0) { 2254 return false; // already found a scale 2255 } 2256 if (n == iv()) { 2257 _scale = 1; 2258 return true; 2259 } 2260 int opc = n->Opcode(); 2261 if (opc == Op_MulI) { 2262 if (n->in(1) == iv() && n->in(2)->is_Con()) { 2263 _scale = n->in(2)->get_int(); 2264 return true; 2265 } else if (n->in(2) == iv() && n->in(1)->is_Con()) { 2266 _scale = n->in(1)->get_int(); 2267 return true; 2268 } 2269 } else if (opc == Op_LShiftI) { 2270 if (n->in(1) == iv() && n->in(2)->is_Con()) { 2271 _scale = 1 << n->in(2)->get_int(); 2272 return true; 2273 } 2274 } else if (opc == Op_ConvI2L) { 2275 if (scaled_iv_plus_offset(n->in(1))) { 2276 return true; 2277 } 2278 } else if (opc == Op_LShiftL) { 2279 if (!has_iv() && _invar == NULL) { 2280 // Need to preserve the current _offset value, so 2281 // create a temporary object for this expression subtree. 2282 // Hacky, so should re-engineer the address pattern match. 2283 SWPointer tmp(this); 2284 if (tmp.scaled_iv_plus_offset(n->in(1))) { 2285 if (tmp._invar == NULL) { 2286 int mult = 1 << n->in(2)->get_int(); 2287 _scale = tmp._scale * mult; 2288 _offset += tmp._offset * mult; 2289 return true; 2290 } 2291 } 2292 } 2293 } 2294 return false; 2295 } 2296 2297 //----------------------------offset_plus_k------------------------ 2298 // Match: offset is (k [+/- invariant]) 2299 // where k maybe zero and invariant is optional, but not both. 2300 bool SWPointer::offset_plus_k(Node* n, bool negate) { 2301 int opc = n->Opcode(); 2302 if (opc == Op_ConI) { 2303 _offset += negate ? -(n->get_int()) : n->get_int(); 2304 return true; 2305 } else if (opc == Op_ConL) { 2306 // Okay if value fits into an int 2307 const TypeLong* t = n->find_long_type(); 2308 if (t->higher_equal(TypeLong::INT)) { 2309 jlong loff = n->get_long(); 2310 jint off = (jint)loff; 2311 _offset += negate ? -off : loff; 2312 return true; 2313 } 2314 return false; 2315 } 2316 if (_invar != NULL) return false; // already have an invariant 2317 if (opc == Op_AddI) { 2318 if (n->in(2)->is_Con() && invariant(n->in(1))) { 2319 _negate_invar = negate; 2320 _invar = n->in(1); 2321 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int(); 2322 return true; 2323 } else if (n->in(1)->is_Con() && invariant(n->in(2))) { 2324 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int(); 2325 _negate_invar = negate; 2326 _invar = n->in(2); 2327 return true; 2328 } 2329 } 2330 if (opc == Op_SubI) { 2331 if (n->in(2)->is_Con() && invariant(n->in(1))) { 2332 _negate_invar = negate; 2333 _invar = n->in(1); 2334 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int(); 2335 return true; 2336 } else if (n->in(1)->is_Con() && invariant(n->in(2))) { 2337 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int(); 2338 _negate_invar = !negate; 2339 _invar = n->in(2); 2340 return true; 2341 } 2342 } 2343 if (invariant(n)) { 2344 _negate_invar = negate; 2345 _invar = n; 2346 return true; 2347 } 2348 return false; 2349 } 2350 2351 //----------------------------print------------------------ 2352 void SWPointer::print() { 2353 #ifndef PRODUCT 2354 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n", 2355 _base != NULL ? _base->_idx : 0, 2356 _adr != NULL ? _adr->_idx : 0, 2357 _scale, _offset, 2358 _negate_invar?'-':'+', 2359 _invar != NULL ? _invar->_idx : 0); 2360 #endif 2361 } 2362 2363 // ========================= OrderedPair ===================== 2364 2365 const OrderedPair OrderedPair::initial; 2366 2367 // ========================= SWNodeInfo ===================== 2368 2369 const SWNodeInfo SWNodeInfo::initial; 2370 2371 2372 // ============================ DepGraph =========================== 2373 2374 //------------------------------make_node--------------------------- 2375 // Make a new dependence graph node for an ideal node. 2376 DepMem* DepGraph::make_node(Node* node) { 2377 DepMem* m = new (_arena) DepMem(node); 2378 if (node != NULL) { 2379 assert(_map.at_grow(node->_idx) == NULL, "one init only"); 2380 _map.at_put_grow(node->_idx, m); 2381 } 2382 return m; 2383 } 2384 2385 //------------------------------make_edge--------------------------- 2386 // Make a new dependence graph edge from dpred -> dsucc 2387 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) { 2388 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head()); 2389 dpred->set_out_head(e); 2390 dsucc->set_in_head(e); 2391 return e; 2392 } 2393 2394 // ========================== DepMem ======================== 2395 2396 //------------------------------in_cnt--------------------------- 2397 int DepMem::in_cnt() { 2398 int ct = 0; 2399 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++; 2400 return ct; 2401 } 2402 2403 //------------------------------out_cnt--------------------------- 2404 int DepMem::out_cnt() { 2405 int ct = 0; 2406 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++; 2407 return ct; 2408 } 2409 2410 //------------------------------print----------------------------- 2411 void DepMem::print() { 2412 #ifndef PRODUCT 2413 tty->print(" DepNode %d (", _node->_idx); 2414 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) { 2415 Node* pred = p->pred()->node(); 2416 tty->print(" %d", pred != NULL ? pred->_idx : 0); 2417 } 2418 tty->print(") ["); 2419 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) { 2420 Node* succ = s->succ()->node(); 2421 tty->print(" %d", succ != NULL ? succ->_idx : 0); 2422 } 2423 tty->print_cr(" ]"); 2424 #endif 2425 } 2426 2427 // =========================== DepEdge ========================= 2428 2429 //------------------------------DepPreds--------------------------- 2430 void DepEdge::print() { 2431 #ifndef PRODUCT 2432 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx); 2433 #endif 2434 } 2435 2436 // =========================== DepPreds ========================= 2437 // Iterator over predecessor edges in the dependence graph. 2438 2439 //------------------------------DepPreds--------------------------- 2440 DepPreds::DepPreds(Node* n, DepGraph& dg) { 2441 _n = n; 2442 _done = false; 2443 if (_n->is_Store() || _n->is_Load()) { 2444 _next_idx = MemNode::Address; 2445 _end_idx = n->req(); 2446 _dep_next = dg.dep(_n)->in_head(); 2447 } else if (_n->is_Mem()) { 2448 _next_idx = 0; 2449 _end_idx = 0; 2450 _dep_next = dg.dep(_n)->in_head(); 2451 } else { 2452 _next_idx = 1; 2453 _end_idx = _n->req(); 2454 _dep_next = NULL; 2455 } 2456 next(); 2457 } 2458 2459 //------------------------------next--------------------------- 2460 void DepPreds::next() { 2461 if (_dep_next != NULL) { 2462 _current = _dep_next->pred()->node(); 2463 _dep_next = _dep_next->next_in(); 2464 } else if (_next_idx < _end_idx) { 2465 _current = _n->in(_next_idx++); 2466 } else { 2467 _done = true; 2468 } 2469 } 2470 2471 // =========================== DepSuccs ========================= 2472 // Iterator over successor edges in the dependence graph. 2473 2474 //------------------------------DepSuccs--------------------------- 2475 DepSuccs::DepSuccs(Node* n, DepGraph& dg) { 2476 _n = n; 2477 _done = false; 2478 if (_n->is_Load()) { 2479 _next_idx = 0; 2480 _end_idx = _n->outcnt(); 2481 _dep_next = dg.dep(_n)->out_head(); 2482 } else if (_n->is_Mem() || _n->is_Phi() && _n->bottom_type() == Type::MEMORY) { 2483 _next_idx = 0; 2484 _end_idx = 0; 2485 _dep_next = dg.dep(_n)->out_head(); 2486 } else { 2487 _next_idx = 0; 2488 _end_idx = _n->outcnt(); 2489 _dep_next = NULL; 2490 } 2491 next(); 2492 } 2493 2494 //-------------------------------next--------------------------- 2495 void DepSuccs::next() { 2496 if (_dep_next != NULL) { 2497 _current = _dep_next->succ()->node(); 2498 _dep_next = _dep_next->next_out(); 2499 } else if (_next_idx < _end_idx) { 2500 _current = _n->raw_out(_next_idx++); 2501 } else { 2502 _done = true; 2503 } 2504 }