1 /* 2 * Copyright (c) 2007, 2018, 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 "memory/resourceArea.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/castnode.hpp" 32 #include "opto/convertnode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/matcher.hpp" 35 #include "opto/memnode.hpp" 36 #include "opto/mulnode.hpp" 37 #include "opto/opcodes.hpp" 38 #include "opto/opaquenode.hpp" 39 #include "opto/superword.hpp" 40 #include "opto/vectornode.hpp" 41 #include "opto/movenode.hpp" 42 43 // 44 // S U P E R W O R D T R A N S F O R M 45 //============================================================================= 46 47 //------------------------------SuperWord--------------------------- 48 SuperWord::SuperWord(PhaseIdealLoop* phase) : 49 _phase(phase), 50 _arena(phase->C->comp_arena()), 51 _igvn(phase->_igvn), 52 _packset(arena(), 8, 0, NULL), // packs for the current block 53 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb 54 _block(arena(), 8, 0, NULL), // nodes in current block 55 _post_block(arena(), 8, 0, NULL), // nodes common to current block which are marked as post loop vectorizable 56 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside 57 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads 58 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails 59 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node 60 _clone_map(phase->C->clone_map()), // map of nodes created in cloning 61 _cmovev_kit(_arena, this), // map to facilitate CMoveV creation 62 _align_to_ref(NULL), // memory reference to align vectors to 63 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs 64 _dg(_arena), // dependence graph 65 _visited(arena()), // visited node set 66 _post_visited(arena()), // post visited node set 67 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs 68 _nlist(arena(), 8, 0, NULL), // scratch list of nodes 69 _stk(arena(), 8, 0, NULL), // scratch stack of nodes 70 _lpt(NULL), // loop tree node 71 _lp(NULL), // LoopNode 72 _bb(NULL), // basic block 73 _iv(NULL), // induction var 74 _race_possible(false), // cases where SDMU is true 75 _early_return(true), // analysis evaluations routine 76 _do_vector_loop(phase->C->do_vector_loop()), // whether to do vectorization/simd style 77 _do_reserve_copy(DoReserveCopyInSuperWord), 78 _num_work_vecs(0), // amount of vector work we have 79 _num_reductions(0), // amount of reduction work we have 80 _ii_first(-1), // first loop generation index - only if do_vector_loop() 81 _ii_last(-1), // last loop generation index - only if do_vector_loop() 82 _ii_order(arena(), 8, 0, 0) 83 { 84 #ifndef PRODUCT 85 _vector_loop_debug = 0; 86 if (_phase->C->method() != NULL) { 87 _vector_loop_debug = phase->C->directive()->VectorizeDebugOption; 88 } 89 90 #endif 91 } 92 93 //------------------------------transform_loop--------------------------- 94 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) { 95 assert(UseSuperWord, "should be"); 96 // Do vectors exist on this architecture? 97 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return; 98 99 assert(lpt->_head->is_CountedLoop(), "must be"); 100 CountedLoopNode *cl = lpt->_head->as_CountedLoop(); 101 102 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop 103 104 bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()); 105 if (post_loop_allowed) { 106 if (cl->is_reduction_loop()) return; // no predication mapping 107 Node *limit = cl->limit(); 108 if (limit->is_Con()) return; // non constant limits only 109 // Now check the limit for expressions we do not handle 110 if (limit->is_Add()) { 111 Node *in2 = limit->in(2); 112 if (in2->is_Con()) { 113 int val = in2->get_int(); 114 // should not try to program these cases 115 if (val < 0) return; 116 } 117 } 118 } 119 120 // skip any loop that has not been assigned max unroll by analysis 121 if (do_optimization) { 122 if (SuperWordLoopUnrollAnalysis && cl->slp_max_unroll() == 0) return; 123 } 124 125 // Check for no control flow in body (other than exit) 126 Node *cl_exit = cl->loopexit(); 127 if (cl->is_main_loop() && (cl_exit->in(0) != lpt->_head)) { 128 #ifndef PRODUCT 129 if (TraceSuperWord) { 130 tty->print_cr("SuperWord::transform_loop: loop too complicated, cl_exit->in(0) != lpt->_head"); 131 tty->print("cl_exit %d", cl_exit->_idx); cl_exit->dump(); 132 tty->print("cl_exit->in(0) %d", cl_exit->in(0)->_idx); cl_exit->in(0)->dump(); 133 tty->print("lpt->_head %d", lpt->_head->_idx); lpt->_head->dump(); 134 lpt->dump_head(); 135 } 136 #endif 137 return; 138 } 139 140 // Make sure the are no extra control users of the loop backedge 141 if (cl->back_control()->outcnt() != 1) { 142 return; 143 } 144 145 // Skip any loops already optimized by slp 146 if (cl->is_vectorized_loop()) return; 147 148 if (cl->do_unroll_only()) return; 149 150 if (cl->is_main_loop()) { 151 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit)))) 152 CountedLoopEndNode* pre_end = get_pre_loop_end(cl); 153 if (pre_end == NULL) return; 154 Node *pre_opaq1 = pre_end->limit(); 155 if (pre_opaq1->Opcode() != Op_Opaque1) return; 156 } 157 158 init(); // initialize data structures 159 160 set_lpt(lpt); 161 set_lp(cl); 162 163 // For now, define one block which is the entire loop body 164 set_bb(cl); 165 166 if (do_optimization) { 167 assert(_packset.length() == 0, "packset must be empty"); 168 SLP_extract(); 169 if (PostLoopMultiversioning && Matcher::has_predicated_vectors()) { 170 if (cl->is_vectorized_loop() && cl->is_main_loop() && !cl->is_reduction_loop()) { 171 IdealLoopTree *lpt_next = lpt->_next; 172 CountedLoopNode *cl_next = lpt_next->_head->as_CountedLoop(); 173 _phase->has_range_checks(lpt_next); 174 if (cl_next->is_post_loop() && !cl_next->range_checks_present()) { 175 if (!cl_next->is_vectorized_loop()) { 176 int slp_max_unroll_factor = cl->slp_max_unroll(); 177 cl_next->set_slp_max_unroll(slp_max_unroll_factor); 178 } 179 } 180 } 181 } 182 } 183 } 184 185 //------------------------------early unrolling analysis------------------------------ 186 void SuperWord::unrolling_analysis(int &local_loop_unroll_factor) { 187 bool is_slp = true; 188 ResourceMark rm; 189 size_t ignored_size = lpt()->_body.size(); 190 int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size); 191 Node_Stack nstack((int)ignored_size); 192 CountedLoopNode *cl = lpt()->_head->as_CountedLoop(); 193 Node *cl_exit = cl->loopexit_or_null(); 194 int rpo_idx = _post_block.length(); 195 196 assert(rpo_idx == 0, "post loop block is empty"); 197 198 // First clear the entries 199 for (uint i = 0; i < lpt()->_body.size(); i++) { 200 ignored_loop_nodes[i] = -1; 201 } 202 203 int max_vector = Matcher::max_vector_size(T_BYTE); 204 bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()); 205 206 // Process the loop, some/all of the stack entries will not be in order, ergo 207 // need to preprocess the ignored initial state before we process the loop 208 for (uint i = 0; i < lpt()->_body.size(); i++) { 209 Node* n = lpt()->_body.at(i); 210 if (n == cl->incr() || 211 n->is_reduction() || 212 n->is_AddP() || 213 n->is_Cmp() || 214 n->is_IfTrue() || 215 n->is_CountedLoop() || 216 (n == cl_exit)) { 217 ignored_loop_nodes[i] = n->_idx; 218 continue; 219 } 220 221 if (n->is_If()) { 222 IfNode *iff = n->as_If(); 223 if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) { 224 if (lpt()->is_loop_exit(iff)) { 225 ignored_loop_nodes[i] = n->_idx; 226 continue; 227 } 228 } 229 } 230 231 if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) { 232 Node* n_tail = n->in(LoopNode::LoopBackControl); 233 if (n_tail != n->in(LoopNode::EntryControl)) { 234 if (!n_tail->is_Mem()) { 235 is_slp = false; 236 break; 237 } 238 } 239 } 240 241 // This must happen after check of phi/if 242 if (n->is_Phi() || n->is_If()) { 243 ignored_loop_nodes[i] = n->_idx; 244 continue; 245 } 246 247 if (n->is_LoadStore() || n->is_MergeMem() || 248 (n->is_Proj() && !n->as_Proj()->is_CFG())) { 249 is_slp = false; 250 break; 251 } 252 253 // Ignore nodes with non-primitive type. 254 BasicType bt; 255 if (n->is_Mem()) { 256 bt = n->as_Mem()->memory_type(); 257 } else { 258 bt = n->bottom_type()->basic_type(); 259 } 260 if (is_java_primitive(bt) == false) { 261 ignored_loop_nodes[i] = n->_idx; 262 continue; 263 } 264 265 if (n->is_Mem()) { 266 MemNode* current = n->as_Mem(); 267 Node* adr = n->in(MemNode::Address); 268 Node* n_ctrl = _phase->get_ctrl(adr); 269 270 // save a queue of post process nodes 271 if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) { 272 // Process the memory expression 273 int stack_idx = 0; 274 bool have_side_effects = true; 275 if (adr->is_AddP() == false) { 276 nstack.push(adr, stack_idx++); 277 } else { 278 // Mark the components of the memory operation in nstack 279 SWPointer p1(current, this, &nstack, true); 280 have_side_effects = p1.node_stack()->is_nonempty(); 281 } 282 283 // Process the pointer stack 284 while (have_side_effects) { 285 Node* pointer_node = nstack.node(); 286 for (uint j = 0; j < lpt()->_body.size(); j++) { 287 Node* cur_node = lpt()->_body.at(j); 288 if (cur_node == pointer_node) { 289 ignored_loop_nodes[j] = cur_node->_idx; 290 break; 291 } 292 } 293 nstack.pop(); 294 have_side_effects = nstack.is_nonempty(); 295 } 296 } 297 } 298 } 299 300 if (is_slp) { 301 // Now we try to find the maximum supported consistent vector which the machine 302 // description can use 303 bool small_basic_type = false; 304 bool flag_small_bt = false; 305 for (uint i = 0; i < lpt()->_body.size(); i++) { 306 if (ignored_loop_nodes[i] != -1) continue; 307 308 BasicType bt; 309 Node* n = lpt()->_body.at(i); 310 if (n->is_Mem()) { 311 bt = n->as_Mem()->memory_type(); 312 } else { 313 bt = n->bottom_type()->basic_type(); 314 } 315 316 if (post_loop_allowed) { 317 if (!small_basic_type) { 318 switch (bt) { 319 case T_CHAR: 320 case T_BYTE: 321 case T_SHORT: 322 small_basic_type = true; 323 break; 324 325 case T_LONG: 326 // TODO: Remove when support completed for mask context with LONG. 327 // Support needs to be augmented for logical qword operations, currently we map to dword 328 // buckets for vectors on logicals as these were legacy. 329 small_basic_type = true; 330 break; 331 332 default: 333 break; 334 } 335 } 336 } 337 338 if (is_java_primitive(bt) == false) continue; 339 340 int cur_max_vector = Matcher::max_vector_size(bt); 341 342 // If a max vector exists which is not larger than _local_loop_unroll_factor 343 // stop looking, we already have the max vector to map to. 344 if (cur_max_vector < local_loop_unroll_factor) { 345 is_slp = false; 346 if (TraceSuperWordLoopUnrollAnalysis) { 347 tty->print_cr("slp analysis fails: unroll limit greater than max vector\n"); 348 } 349 break; 350 } 351 352 // Map the maximal common vector 353 if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) { 354 if (cur_max_vector < max_vector && !flag_small_bt) { 355 max_vector = cur_max_vector; 356 } else if (cur_max_vector > max_vector && UseSubwordForMaxVector) { 357 // Analyse subword in the loop to set maximum vector size to take advantage of full vector width for subword types. 358 // Here we analyze if narrowing is likely to happen and if it is we set vector size more aggressively. 359 // We check for possibility of narrowing by looking through chain operations using subword types. 360 if (is_subword_type(bt)) { 361 uint start, end; 362 VectorNode::vector_operands(n, &start, &end); 363 364 for (uint j = start; j < end; j++) { 365 Node* in = n->in(j); 366 // Don't propagate through a memory 367 if (!in->is_Mem() && in_bb(in) && in->bottom_type()->basic_type() == T_INT) { 368 bool same_type = true; 369 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) { 370 Node *use = in->fast_out(k); 371 if (!in_bb(use) && use->bottom_type()->basic_type() != bt) { 372 same_type = false; 373 break; 374 } 375 } 376 if (same_type) { 377 max_vector = cur_max_vector; 378 flag_small_bt = true; 379 cl->mark_subword_loop(); 380 } 381 } 382 } 383 } 384 } 385 // We only process post loops on predicated targets where we want to 386 // mask map the loop to a single iteration 387 if (post_loop_allowed) { 388 _post_block.at_put_grow(rpo_idx++, n); 389 } 390 } 391 } 392 if (is_slp) { 393 local_loop_unroll_factor = max_vector; 394 cl->mark_passed_slp(); 395 } 396 cl->mark_was_slp(); 397 if (cl->is_main_loop()) { 398 cl->set_slp_max_unroll(local_loop_unroll_factor); 399 } else if (post_loop_allowed) { 400 if (!small_basic_type) { 401 // avoid replication context for small basic types in programmable masked loops 402 cl->set_slp_max_unroll(local_loop_unroll_factor); 403 } 404 } 405 } 406 } 407 408 //------------------------------SLP_extract--------------------------- 409 // Extract the superword level parallelism 410 // 411 // 1) A reverse post-order of nodes in the block is constructed. By scanning 412 // this list from first to last, all definitions are visited before their uses. 413 // 414 // 2) A point-to-point dependence graph is constructed between memory references. 415 // This simplies the upcoming "independence" checker. 416 // 417 // 3) The maximum depth in the node graph from the beginning of the block 418 // to each node is computed. This is used to prune the graph search 419 // in the independence checker. 420 // 421 // 4) For integer types, the necessary bit width is propagated backwards 422 // from stores to allow packed operations on byte, char, and short 423 // integers. This reverses the promotion to type "int" that javac 424 // did for operations like: char c1,c2,c3; c1 = c2 + c3. 425 // 426 // 5) One of the memory references is picked to be an aligned vector reference. 427 // The pre-loop trip count is adjusted to align this reference in the 428 // unrolled body. 429 // 430 // 6) The initial set of pack pairs is seeded with memory references. 431 // 432 // 7) The set of pack pairs is extended by following use->def and def->use links. 433 // 434 // 8) The pairs are combined into vector sized packs. 435 // 436 // 9) Reorder the memory slices to co-locate members of the memory packs. 437 // 438 // 10) Generate ideal vector nodes for the final set of packs and where necessary, 439 // inserting scalar promotion, vector creation from multiple scalars, and 440 // extraction of scalar values from vectors. 441 // 442 void SuperWord::SLP_extract() { 443 444 #ifndef PRODUCT 445 if (_do_vector_loop && TraceSuperWord) { 446 tty->print("SuperWord::SLP_extract\n"); 447 tty->print("input loop\n"); 448 _lpt->dump_head(); 449 _lpt->dump(); 450 for (uint i = 0; i < _lpt->_body.size(); i++) { 451 _lpt->_body.at(i)->dump(); 452 } 453 } 454 #endif 455 // Ready the block 456 if (!construct_bb()) { 457 return; // Exit if no interesting nodes or complex graph. 458 } 459 460 // build _dg, _disjoint_ptrs 461 dependence_graph(); 462 463 // compute function depth(Node*) 464 compute_max_depth(); 465 466 CountedLoopNode *cl = lpt()->_head->as_CountedLoop(); 467 bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()); 468 if (cl->is_main_loop()) { 469 if (_do_vector_loop) { 470 if (mark_generations() != -1) { 471 hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly 472 473 if (!construct_bb()) { 474 return; // Exit if no interesting nodes or complex graph. 475 } 476 dependence_graph(); 477 compute_max_depth(); 478 } 479 480 #ifndef PRODUCT 481 if (TraceSuperWord) { 482 tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph"); 483 _lpt->dump_head(); 484 for (int j = 0; j < _block.length(); j++) { 485 Node* n = _block.at(j); 486 int d = depth(n); 487 for (int i = 0; i < d; i++) tty->print("%s", " "); 488 tty->print("%d :", d); 489 n->dump(); 490 } 491 } 492 #endif 493 } 494 495 compute_vector_element_type(); 496 497 // Attempt vectorization 498 499 find_adjacent_refs(); 500 501 extend_packlist(); 502 503 if (_do_vector_loop) { 504 if (_packset.length() == 0) { 505 if (TraceSuperWord) { 506 tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway"); 507 } 508 pack_parallel(); 509 } 510 } 511 512 combine_packs(); 513 514 construct_my_pack_map(); 515 if (UseVectorCmov) { 516 merge_packs_to_cmovd(); 517 } 518 519 filter_packs(); 520 521 schedule(); 522 } else if (post_loop_allowed) { 523 int saved_mapped_unroll_factor = cl->slp_max_unroll(); 524 if (saved_mapped_unroll_factor) { 525 int vector_mapped_unroll_factor = saved_mapped_unroll_factor; 526 527 // now reset the slp_unroll_factor so that we can check the analysis mapped 528 // what the vector loop was mapped to 529 cl->set_slp_max_unroll(0); 530 531 // do the analysis on the post loop 532 unrolling_analysis(vector_mapped_unroll_factor); 533 534 // if our analyzed loop is a canonical fit, start processing it 535 if (vector_mapped_unroll_factor == saved_mapped_unroll_factor) { 536 // now add the vector nodes to packsets 537 for (int i = 0; i < _post_block.length(); i++) { 538 Node* n = _post_block.at(i); 539 Node_List* singleton = new Node_List(); 540 singleton->push(n); 541 _packset.append(singleton); 542 set_my_pack(n, singleton); 543 } 544 545 // map base types for vector usage 546 compute_vector_element_type(); 547 } else { 548 return; 549 } 550 } else { 551 // for some reason we could not map the slp analysis state of the vectorized loop 552 return; 553 } 554 } 555 556 output(); 557 } 558 559 //------------------------------find_adjacent_refs--------------------------- 560 // Find the adjacent memory references and create pack pairs for them. 561 // This is the initial set of packs that will then be extended by 562 // following use->def and def->use links. The align positions are 563 // assigned relative to the reference "align_to_ref" 564 void SuperWord::find_adjacent_refs() { 565 // Get list of memory operations 566 Node_List memops; 567 for (int i = 0; i < _block.length(); i++) { 568 Node* n = _block.at(i); 569 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) && 570 is_java_primitive(n->as_Mem()->memory_type())) { 571 int align = memory_alignment(n->as_Mem(), 0); 572 if (align != bottom_align) { 573 memops.push(n); 574 } 575 } 576 } 577 578 Node_List align_to_refs; 579 int best_iv_adjustment = 0; 580 MemNode* best_align_to_mem_ref = NULL; 581 582 while (memops.size() != 0) { 583 // Find a memory reference to align to. 584 MemNode* mem_ref = find_align_to_ref(memops); 585 if (mem_ref == NULL) break; 586 align_to_refs.push(mem_ref); 587 int iv_adjustment = get_iv_adjustment(mem_ref); 588 589 if (best_align_to_mem_ref == NULL) { 590 // Set memory reference which is the best from all memory operations 591 // to be used for alignment. The pre-loop trip count is modified to align 592 // this reference to a vector-aligned address. 593 best_align_to_mem_ref = mem_ref; 594 best_iv_adjustment = iv_adjustment; 595 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);) 596 } 597 598 SWPointer align_to_ref_p(mem_ref, this, NULL, false); 599 // Set alignment relative to "align_to_ref" for all related memory operations. 600 for (int i = memops.size() - 1; i >= 0; i--) { 601 MemNode* s = memops.at(i)->as_Mem(); 602 if (isomorphic(s, mem_ref) && 603 (!_do_vector_loop || same_origin_idx(s, mem_ref))) { 604 SWPointer p2(s, this, NULL, false); 605 if (p2.comparable(align_to_ref_p)) { 606 int align = memory_alignment(s, iv_adjustment); 607 set_alignment(s, align); 608 } 609 } 610 } 611 612 // Create initial pack pairs of memory operations for which 613 // alignment is set and vectors will be aligned. 614 bool create_pack = true; 615 if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) { 616 if (!Matcher::misaligned_vectors_ok()) { 617 int vw = vector_width(mem_ref); 618 int vw_best = vector_width(best_align_to_mem_ref); 619 if (vw > vw_best) { 620 // Do not vectorize a memory access with more elements per vector 621 // if unaligned memory access is not allowed because number of 622 // iterations in pre-loop will be not enough to align it. 623 create_pack = false; 624 } else { 625 SWPointer p2(best_align_to_mem_ref, this, NULL, false); 626 if (align_to_ref_p.invar() != p2.invar()) { 627 // Do not vectorize memory accesses with different invariants 628 // if unaligned memory accesses are not allowed. 629 create_pack = false; 630 } 631 } 632 } 633 } else { 634 if (same_velt_type(mem_ref, best_align_to_mem_ref)) { 635 // Can't allow vectorization of unaligned memory accesses with the 636 // same type since it could be overlapped accesses to the same array. 637 create_pack = false; 638 } else { 639 // Allow independent (different type) unaligned memory operations 640 // if HW supports them. 641 if (!Matcher::misaligned_vectors_ok()) { 642 create_pack = false; 643 } else { 644 // Check if packs of the same memory type but 645 // with a different alignment were created before. 646 for (uint i = 0; i < align_to_refs.size(); i++) { 647 MemNode* mr = align_to_refs.at(i)->as_Mem(); 648 if (mr == mem_ref) { 649 // Skip when we are looking at same memory operation. 650 continue; 651 } 652 if (same_velt_type(mr, mem_ref) && 653 memory_alignment(mr, iv_adjustment) != 0) 654 create_pack = false; 655 } 656 } 657 } 658 } 659 if (create_pack) { 660 for (uint i = 0; i < memops.size(); i++) { 661 Node* s1 = memops.at(i); 662 int align = alignment(s1); 663 if (align == top_align) continue; 664 for (uint j = 0; j < memops.size(); j++) { 665 Node* s2 = memops.at(j); 666 if (alignment(s2) == top_align) continue; 667 if (s1 != s2 && are_adjacent_refs(s1, s2)) { 668 if (stmts_can_pack(s1, s2, align)) { 669 Node_List* pair = new Node_List(); 670 pair->push(s1); 671 pair->push(s2); 672 if (!_do_vector_loop || same_origin_idx(s1, s2)) { 673 _packset.append(pair); 674 } 675 } 676 } 677 } 678 } 679 } else { // Don't create unaligned pack 680 // First, remove remaining memory ops of the same type from the list. 681 for (int i = memops.size() - 1; i >= 0; i--) { 682 MemNode* s = memops.at(i)->as_Mem(); 683 if (same_velt_type(s, mem_ref)) { 684 memops.remove(i); 685 } 686 } 687 688 // Second, remove already constructed packs of the same type. 689 for (int i = _packset.length() - 1; i >= 0; i--) { 690 Node_List* p = _packset.at(i); 691 MemNode* s = p->at(0)->as_Mem(); 692 if (same_velt_type(s, mem_ref)) { 693 remove_pack_at(i); 694 } 695 } 696 697 // If needed find the best memory reference for loop alignment again. 698 if (same_velt_type(mem_ref, best_align_to_mem_ref)) { 699 // Put memory ops from remaining packs back on memops list for 700 // the best alignment search. 701 uint orig_msize = memops.size(); 702 for (int i = 0; i < _packset.length(); i++) { 703 Node_List* p = _packset.at(i); 704 MemNode* s = p->at(0)->as_Mem(); 705 assert(!same_velt_type(s, mem_ref), "sanity"); 706 memops.push(s); 707 } 708 best_align_to_mem_ref = find_align_to_ref(memops); 709 if (best_align_to_mem_ref == NULL) { 710 if (TraceSuperWord) { 711 tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL"); 712 } 713 break; 714 } 715 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref); 716 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);) 717 // Restore list. 718 while (memops.size() > orig_msize) 719 (void)memops.pop(); 720 } 721 } // unaligned memory accesses 722 723 // Remove used mem nodes. 724 for (int i = memops.size() - 1; i >= 0; i--) { 725 MemNode* m = memops.at(i)->as_Mem(); 726 if (alignment(m) != top_align) { 727 memops.remove(i); 728 } 729 } 730 731 } // while (memops.size() != 0 732 set_align_to_ref(best_align_to_mem_ref); 733 734 if (TraceSuperWord) { 735 tty->print_cr("\nAfter find_adjacent_refs"); 736 print_packset(); 737 } 738 } 739 740 #ifndef PRODUCT 741 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) { 742 if (is_trace_adjacent()) { 743 tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d", 744 best_align_to_mem_ref->_idx, best_iv_adjustment); 745 best_align_to_mem_ref->dump(); 746 } 747 } 748 #endif 749 750 //------------------------------find_align_to_ref--------------------------- 751 // Find a memory reference to align the loop induction variable to. 752 // Looks first at stores then at loads, looking for a memory reference 753 // with the largest number of references similar to it. 754 MemNode* SuperWord::find_align_to_ref(Node_List &memops) { 755 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0); 756 757 // Count number of comparable memory ops 758 for (uint i = 0; i < memops.size(); i++) { 759 MemNode* s1 = memops.at(i)->as_Mem(); 760 SWPointer p1(s1, this, NULL, false); 761 // Discard if pre loop can't align this reference 762 if (!ref_is_alignable(p1)) { 763 *cmp_ct.adr_at(i) = 0; 764 continue; 765 } 766 for (uint j = i+1; j < memops.size(); j++) { 767 MemNode* s2 = memops.at(j)->as_Mem(); 768 if (isomorphic(s1, s2)) { 769 SWPointer p2(s2, this, NULL, false); 770 if (p1.comparable(p2)) { 771 (*cmp_ct.adr_at(i))++; 772 (*cmp_ct.adr_at(j))++; 773 } 774 } 775 } 776 } 777 778 // Find Store (or Load) with the greatest number of "comparable" references, 779 // biggest vector size, smallest data size and smallest iv offset. 780 int max_ct = 0; 781 int max_vw = 0; 782 int max_idx = -1; 783 int min_size = max_jint; 784 int min_iv_offset = max_jint; 785 for (uint j = 0; j < memops.size(); j++) { 786 MemNode* s = memops.at(j)->as_Mem(); 787 if (s->is_Store()) { 788 int vw = vector_width_in_bytes(s); 789 assert(vw > 1, "sanity"); 790 SWPointer p(s, this, NULL, false); 791 if ( cmp_ct.at(j) > max_ct || 792 (cmp_ct.at(j) == max_ct && 793 ( vw > max_vw || 794 (vw == max_vw && 795 ( data_size(s) < min_size || 796 (data_size(s) == min_size && 797 p.offset_in_bytes() < min_iv_offset)))))) { 798 max_ct = cmp_ct.at(j); 799 max_vw = vw; 800 max_idx = j; 801 min_size = data_size(s); 802 min_iv_offset = p.offset_in_bytes(); 803 } 804 } 805 } 806 // If no stores, look at loads 807 if (max_ct == 0) { 808 for (uint j = 0; j < memops.size(); j++) { 809 MemNode* s = memops.at(j)->as_Mem(); 810 if (s->is_Load()) { 811 int vw = vector_width_in_bytes(s); 812 assert(vw > 1, "sanity"); 813 SWPointer p(s, this, NULL, false); 814 if ( cmp_ct.at(j) > max_ct || 815 (cmp_ct.at(j) == max_ct && 816 ( vw > max_vw || 817 (vw == max_vw && 818 ( data_size(s) < min_size || 819 (data_size(s) == min_size && 820 p.offset_in_bytes() < min_iv_offset)))))) { 821 max_ct = cmp_ct.at(j); 822 max_vw = vw; 823 max_idx = j; 824 min_size = data_size(s); 825 min_iv_offset = p.offset_in_bytes(); 826 } 827 } 828 } 829 } 830 831 #ifdef ASSERT 832 if (TraceSuperWord && Verbose) { 833 tty->print_cr("\nVector memops after find_align_to_ref"); 834 for (uint i = 0; i < memops.size(); i++) { 835 MemNode* s = memops.at(i)->as_Mem(); 836 s->dump(); 837 } 838 } 839 #endif 840 841 if (max_ct > 0) { 842 #ifdef ASSERT 843 if (TraceSuperWord) { 844 tty->print("\nVector align to node: "); 845 memops.at(max_idx)->as_Mem()->dump(); 846 } 847 #endif 848 return memops.at(max_idx)->as_Mem(); 849 } 850 return NULL; 851 } 852 853 //------------------span_works_for_memory_size----------------------------- 854 static bool span_works_for_memory_size(MemNode* mem, int span, int mem_size, int offset) { 855 bool span_matches_memory = false; 856 if ((mem_size == type2aelembytes(T_BYTE) || mem_size == type2aelembytes(T_SHORT)) 857 && ABS(span) == type2aelembytes(T_INT)) { 858 // There is a mismatch on span size compared to memory. 859 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) { 860 Node* use = mem->fast_out(j); 861 if (!VectorNode::is_type_transition_to_int(use)) { 862 return false; 863 } 864 } 865 // If all uses transition to integer, it means that we can successfully align even on mismatch. 866 return true; 867 } 868 else { 869 span_matches_memory = ABS(span) == mem_size; 870 } 871 return span_matches_memory && (ABS(offset) % mem_size) == 0; 872 } 873 874 //------------------------------ref_is_alignable--------------------------- 875 // Can the preloop align the reference to position zero in the vector? 876 bool SuperWord::ref_is_alignable(SWPointer& p) { 877 if (!p.has_iv()) { 878 return true; // no induction variable 879 } 880 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop()); 881 assert(pre_end != NULL, "we must have a correct pre-loop"); 882 assert(pre_end->stride_is_con(), "pre loop stride is constant"); 883 int preloop_stride = pre_end->stride_con(); 884 885 int span = preloop_stride * p.scale_in_bytes(); 886 int mem_size = p.memory_size(); 887 int offset = p.offset_in_bytes(); 888 // Stride one accesses are alignable if offset is aligned to memory operation size. 889 // Offset can be unaligned when UseUnalignedAccesses is used. 890 if (span_works_for_memory_size(p.mem(), span, mem_size, offset)) { 891 return true; 892 } 893 // If the initial offset from start of the object is computable, 894 // check if the pre-loop can align the final offset accordingly. 895 // 896 // In other words: Can we find an i such that the offset 897 // after i pre-loop iterations is aligned to vw? 898 // (init_offset + pre_loop) % vw == 0 (1) 899 // where 900 // pre_loop = i * span 901 // is the number of bytes added to the offset by i pre-loop iterations. 902 // 903 // For this to hold we need pre_loop to increase init_offset by 904 // pre_loop = vw - (init_offset % vw) 905 // 906 // This is only possible if pre_loop is divisible by span because each 907 // pre-loop iteration increases the initial offset by 'span' bytes: 908 // (vw - (init_offset % vw)) % span == 0 909 // 910 int vw = vector_width_in_bytes(p.mem()); 911 assert(vw > 1, "sanity"); 912 Node* init_nd = pre_end->init_trip(); 913 if (init_nd->is_Con() && p.invar() == NULL) { 914 int init = init_nd->bottom_type()->is_int()->get_con(); 915 int init_offset = init * p.scale_in_bytes() + offset; 916 if (init_offset < 0) { // negative offset from object start? 917 return false; // may happen in dead loop 918 } 919 if (vw % span == 0) { 920 // If vm is a multiple of span, we use formula (1). 921 if (span > 0) { 922 return (vw - (init_offset % vw)) % span == 0; 923 } else { 924 assert(span < 0, "nonzero stride * scale"); 925 return (init_offset % vw) % -span == 0; 926 } 927 } else if (span % vw == 0) { 928 // If span is a multiple of vw, we can simplify formula (1) to: 929 // (init_offset + i * span) % vw == 0 930 // => 931 // (init_offset % vw) + ((i * span) % vw) == 0 932 // => 933 // init_offset % vw == 0 934 // 935 // Because we add a multiple of vw to the initial offset, the final 936 // offset is a multiple of vw if and only if init_offset is a multiple. 937 // 938 return (init_offset % vw) == 0; 939 } 940 } 941 return false; 942 } 943 //---------------------------get_vw_bytes_special------------------------ 944 int SuperWord::get_vw_bytes_special(MemNode* s) { 945 // Get the vector width in bytes. 946 int vw = vector_width_in_bytes(s); 947 948 // Check for special case where there is an MulAddS2I usage where short vectors are going to need combined. 949 BasicType btype = velt_basic_type(s); 950 if (type2aelembytes(btype) == 2) { 951 bool should_combine_adjacent = true; 952 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { 953 Node* user = s->fast_out(i); 954 if (!VectorNode::is_muladds2i(user)) { 955 should_combine_adjacent = false; 956 } 957 } 958 if (should_combine_adjacent) { 959 vw = MIN2(Matcher::max_vector_size(btype)*type2aelembytes(btype), vw * 2); 960 } 961 } 962 963 return vw; 964 } 965 966 //---------------------------get_iv_adjustment--------------------------- 967 // Calculate loop's iv adjustment for this memory ops. 968 int SuperWord::get_iv_adjustment(MemNode* mem_ref) { 969 SWPointer align_to_ref_p(mem_ref, this, NULL, false); 970 int offset = align_to_ref_p.offset_in_bytes(); 971 int scale = align_to_ref_p.scale_in_bytes(); 972 int elt_size = align_to_ref_p.memory_size(); 973 int vw = get_vw_bytes_special(mem_ref); 974 assert(vw > 1, "sanity"); 975 int iv_adjustment; 976 if (scale != 0) { 977 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1; 978 // At least one iteration is executed in pre-loop by default. As result 979 // several iterations are needed to align memory operations in main-loop even 980 // if offset is 0. 981 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw)); 982 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0), 983 "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size); 984 iv_adjustment = iv_adjustment_in_bytes/elt_size; 985 } else { 986 // This memory op is not dependent on iv (scale == 0) 987 iv_adjustment = 0; 988 } 989 990 #ifndef PRODUCT 991 if (TraceSuperWord) { 992 tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ", 993 mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw); 994 mem_ref->dump(); 995 } 996 #endif 997 return iv_adjustment; 998 } 999 1000 //---------------------------dependence_graph--------------------------- 1001 // Construct dependency graph. 1002 // Add dependence edges to load/store nodes for memory dependence 1003 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x) 1004 void SuperWord::dependence_graph() { 1005 CountedLoopNode *cl = lpt()->_head->as_CountedLoop(); 1006 // First, assign a dependence node to each memory node 1007 for (int i = 0; i < _block.length(); i++ ) { 1008 Node *n = _block.at(i); 1009 if (n->is_Mem() || (n->is_Phi() && n->bottom_type() == Type::MEMORY)) { 1010 _dg.make_node(n); 1011 } 1012 } 1013 1014 // For each memory slice, create the dependences 1015 for (int i = 0; i < _mem_slice_head.length(); i++) { 1016 Node* n = _mem_slice_head.at(i); 1017 Node* n_tail = _mem_slice_tail.at(i); 1018 1019 // Get slice in predecessor order (last is first) 1020 if (cl->is_main_loop()) { 1021 mem_slice_preds(n_tail, n, _nlist); 1022 } 1023 1024 #ifndef PRODUCT 1025 if(TraceSuperWord && Verbose) { 1026 tty->print_cr("SuperWord::dependence_graph: built a new mem slice"); 1027 for (int j = _nlist.length() - 1; j >= 0 ; j--) { 1028 _nlist.at(j)->dump(); 1029 } 1030 } 1031 #endif 1032 // Make the slice dependent on the root 1033 DepMem* slice = _dg.dep(n); 1034 _dg.make_edge(_dg.root(), slice); 1035 1036 // Create a sink for the slice 1037 DepMem* slice_sink = _dg.make_node(NULL); 1038 _dg.make_edge(slice_sink, _dg.tail()); 1039 1040 // Now visit each pair of memory ops, creating the edges 1041 for (int j = _nlist.length() - 1; j >= 0 ; j--) { 1042 Node* s1 = _nlist.at(j); 1043 1044 // If no dependency yet, use slice 1045 if (_dg.dep(s1)->in_cnt() == 0) { 1046 _dg.make_edge(slice, s1); 1047 } 1048 SWPointer p1(s1->as_Mem(), this, NULL, false); 1049 bool sink_dependent = true; 1050 for (int k = j - 1; k >= 0; k--) { 1051 Node* s2 = _nlist.at(k); 1052 if (s1->is_Load() && s2->is_Load()) 1053 continue; 1054 SWPointer p2(s2->as_Mem(), this, NULL, false); 1055 1056 int cmp = p1.cmp(p2); 1057 if (SuperWordRTDepCheck && 1058 p1.base() != p2.base() && p1.valid() && p2.valid()) { 1059 // Create a runtime check to disambiguate 1060 OrderedPair pp(p1.base(), p2.base()); 1061 _disjoint_ptrs.append_if_missing(pp); 1062 } else if (!SWPointer::not_equal(cmp)) { 1063 // Possibly same address 1064 _dg.make_edge(s1, s2); 1065 sink_dependent = false; 1066 } 1067 } 1068 if (sink_dependent) { 1069 _dg.make_edge(s1, slice_sink); 1070 } 1071 } 1072 1073 if (TraceSuperWord) { 1074 tty->print_cr("\nDependence graph for slice: %d", n->_idx); 1075 for (int q = 0; q < _nlist.length(); q++) { 1076 _dg.print(_nlist.at(q)); 1077 } 1078 tty->cr(); 1079 } 1080 1081 _nlist.clear(); 1082 } 1083 1084 if (TraceSuperWord) { 1085 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE"); 1086 for (int r = 0; r < _disjoint_ptrs.length(); r++) { 1087 _disjoint_ptrs.at(r).print(); 1088 tty->cr(); 1089 } 1090 tty->cr(); 1091 } 1092 1093 } 1094 1095 //---------------------------mem_slice_preds--------------------------- 1096 // Return a memory slice (node list) in predecessor order starting at "start" 1097 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) { 1098 assert(preds.length() == 0, "start empty"); 1099 Node* n = start; 1100 Node* prev = NULL; 1101 while (true) { 1102 NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);) 1103 assert(in_bb(n), "must be in block"); 1104 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1105 Node* out = n->fast_out(i); 1106 if (out->is_Load()) { 1107 if (in_bb(out)) { 1108 preds.push(out); 1109 if (TraceSuperWord && Verbose) { 1110 tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx); 1111 } 1112 } 1113 } else { 1114 // FIXME 1115 if (out->is_MergeMem() && !in_bb(out)) { 1116 // Either unrolling is causing a memory edge not to disappear, 1117 // or need to run igvn.optimize() again before SLP 1118 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) { 1119 // Ditto. Not sure what else to check further. 1120 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) { 1121 // StoreCM has an input edge used as a precedence edge. 1122 // Maybe an issue when oop stores are vectorized. 1123 } else { 1124 assert(out == prev || prev == NULL, "no branches off of store slice"); 1125 } 1126 }//else 1127 }//for 1128 if (n == stop) break; 1129 preds.push(n); 1130 if (TraceSuperWord && Verbose) { 1131 tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx); 1132 } 1133 prev = n; 1134 assert(n->is_Mem(), "unexpected node %s", n->Name()); 1135 n = n->in(MemNode::Memory); 1136 } 1137 } 1138 1139 //------------------------------stmts_can_pack--------------------------- 1140 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and 1141 // s1 aligned at "align" 1142 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) { 1143 1144 // Do not use superword for non-primitives 1145 BasicType bt1 = velt_basic_type(s1); 1146 BasicType bt2 = velt_basic_type(s2); 1147 if(!is_java_primitive(bt1) || !is_java_primitive(bt2)) 1148 return false; 1149 if (Matcher::max_vector_size(bt1) < 2) { 1150 return false; // No vectors for this type 1151 } 1152 1153 if (isomorphic(s1, s2)) { 1154 if ((independent(s1, s2) && have_similar_inputs(s1, s2)) || reduction(s1, s2)) { 1155 if (!exists_at(s1, 0) && !exists_at(s2, 1)) { 1156 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) { 1157 int s1_align = alignment(s1); 1158 int s2_align = alignment(s2); 1159 if (s1_align == top_align || s1_align == align) { 1160 if (s2_align == top_align || s2_align == align + data_size(s1)) { 1161 return true; 1162 } 1163 } 1164 } 1165 } 1166 } 1167 } 1168 return false; 1169 } 1170 1171 //------------------------------exists_at--------------------------- 1172 // Does s exist in a pack at position pos? 1173 bool SuperWord::exists_at(Node* s, uint pos) { 1174 for (int i = 0; i < _packset.length(); i++) { 1175 Node_List* p = _packset.at(i); 1176 if (p->at(pos) == s) { 1177 return true; 1178 } 1179 } 1180 return false; 1181 } 1182 1183 //------------------------------are_adjacent_refs--------------------------- 1184 // Is s1 immediately before s2 in memory? 1185 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) { 1186 if (!s1->is_Mem() || !s2->is_Mem()) return false; 1187 if (!in_bb(s1) || !in_bb(s2)) return false; 1188 1189 // Do not use superword for non-primitives 1190 if (!is_java_primitive(s1->as_Mem()->memory_type()) || 1191 !is_java_primitive(s2->as_Mem()->memory_type())) { 1192 return false; 1193 } 1194 1195 // FIXME - co_locate_pack fails on Stores in different mem-slices, so 1196 // only pack memops that are in the same alias set until that's fixed. 1197 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) != 1198 _phase->C->get_alias_index(s2->as_Mem()->adr_type())) 1199 return false; 1200 SWPointer p1(s1->as_Mem(), this, NULL, false); 1201 SWPointer p2(s2->as_Mem(), this, NULL, false); 1202 if (p1.base() != p2.base() || !p1.comparable(p2)) return false; 1203 int diff = p2.offset_in_bytes() - p1.offset_in_bytes(); 1204 return diff == data_size(s1); 1205 } 1206 1207 //------------------------------isomorphic--------------------------- 1208 // Are s1 and s2 similar? 1209 bool SuperWord::isomorphic(Node* s1, Node* s2) { 1210 if (s1->Opcode() != s2->Opcode()) return false; 1211 if (s1->req() != s2->req()) return false; 1212 if (!same_velt_type(s1, s2)) return false; 1213 Node* s1_ctrl = s1->in(0); 1214 Node* s2_ctrl = s2->in(0); 1215 // If the control nodes are equivalent, no further checks are required to test for isomorphism. 1216 if (s1_ctrl == s2_ctrl) { 1217 return true; 1218 } else { 1219 bool ret = false; 1220 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { 1221 Node* t1 = s1->fast_out(i); 1222 for (DUIterator_Fast imax, i = s2->fast_outs(imax); i < imax; i++) { 1223 Node* t2 = s2->fast_out(i); 1224 if (VectorNode::is_muladds2i(t1) && VectorNode::is_muladds2i(t2)) { 1225 bool s1_ctrl_inv = ((s1_ctrl == NULL) ? true : lpt()->is_invariant(s1_ctrl)); 1226 bool s2_ctrl_inv = ((s2_ctrl == NULL) ? true : lpt()->is_invariant(s2_ctrl)); 1227 // If the control nodes are not invariant for the loop, fail isomorphism test. 1228 if (!s1_ctrl_inv || !s2_ctrl_inv) { 1229 return false; 1230 } 1231 if (s1_ctrl->is_Proj()) { 1232 s1_ctrl = s1_ctrl->in(0); 1233 assert(lpt()->is_invariant(s1_ctrl), "must be invariant"); 1234 } 1235 if (s2_ctrl->is_Proj()) { 1236 s2_ctrl = s2_ctrl->in(0); 1237 assert(lpt()->is_invariant(s2_ctrl), "must be invariant"); 1238 } 1239 // Control nodes are invariant. However, we have no way of checking whether they resolve 1240 // in an equivalent manner. But, we know that invariant range checks are guaranteed to 1241 // throw before the loop (if they would have thrown). Thus, the loop would not have been reached. 1242 // Therefore, if the control nodes for both are range checks, we accept them to be isomorphic. 1243 if (!s1_ctrl->is_RangeCheck() || !s2_ctrl->is_RangeCheck()) { 1244 return false; 1245 } 1246 ret = true; 1247 } 1248 } 1249 } 1250 return ret; 1251 } 1252 } 1253 1254 //------------------------------independent--------------------------- 1255 // Is there no data path from s1 to s2 or s2 to s1? 1256 bool SuperWord::independent(Node* s1, Node* s2) { 1257 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first"); 1258 int d1 = depth(s1); 1259 int d2 = depth(s2); 1260 if (d1 == d2) return s1 != s2; 1261 Node* deep = d1 > d2 ? s1 : s2; 1262 Node* shallow = d1 > d2 ? s2 : s1; 1263 1264 visited_clear(); 1265 1266 return independent_path(shallow, deep); 1267 } 1268 1269 //--------------------------have_similar_inputs----------------------- 1270 // For a node pair (s1, s2) which is isomorphic and independent, 1271 // do s1 and s2 have similar input edges? 1272 bool SuperWord::have_similar_inputs(Node* s1, Node* s2) { 1273 // assert(isomorphic(s1, s2) == true, "check isomorphic"); 1274 // assert(independent(s1, s2) == true, "check independent"); 1275 if (s1->req() > 1 && !s1->is_Store() && !s1->is_Load()) { 1276 for (uint i = 1; i < s1->req(); i++) { 1277 if (s1->in(i)->Opcode() != s2->in(i)->Opcode()) return false; 1278 } 1279 } 1280 return true; 1281 } 1282 1283 //------------------------------reduction--------------------------- 1284 // Is there a data path between s1 and s2 and the nodes reductions? 1285 bool SuperWord::reduction(Node* s1, Node* s2) { 1286 bool retValue = false; 1287 int d1 = depth(s1); 1288 int d2 = depth(s2); 1289 if (d1 + 1 == d2) { 1290 if (s1->is_reduction() && s2->is_reduction()) { 1291 // This is an ordered set, so s1 should define s2 1292 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { 1293 Node* t1 = s1->fast_out(i); 1294 if (t1 == s2) { 1295 // both nodes are reductions and connected 1296 retValue = true; 1297 } 1298 } 1299 } 1300 } 1301 1302 return retValue; 1303 } 1304 1305 //------------------------------independent_path------------------------------ 1306 // Helper for independent 1307 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) { 1308 if (dp >= 1000) return false; // stop deep recursion 1309 visited_set(deep); 1310 int shal_depth = depth(shallow); 1311 assert(shal_depth <= depth(deep), "must be"); 1312 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) { 1313 Node* pred = preds.current(); 1314 if (in_bb(pred) && !visited_test(pred)) { 1315 if (shallow == pred) { 1316 return false; 1317 } 1318 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) { 1319 return false; 1320 } 1321 } 1322 } 1323 return true; 1324 } 1325 1326 //------------------------------set_alignment--------------------------- 1327 void SuperWord::set_alignment(Node* s1, Node* s2, int align) { 1328 set_alignment(s1, align); 1329 if (align == top_align || align == bottom_align) { 1330 set_alignment(s2, align); 1331 } else { 1332 set_alignment(s2, align + data_size(s1)); 1333 } 1334 } 1335 1336 //------------------------------data_size--------------------------- 1337 int SuperWord::data_size(Node* s) { 1338 Node* use = NULL; //test if the node is a candidate for CMoveV optimization, then return the size of CMov 1339 if (UseVectorCmov) { 1340 use = _cmovev_kit.is_Bool_candidate(s); 1341 if (use != NULL) { 1342 return data_size(use); 1343 } 1344 use = _cmovev_kit.is_CmpD_candidate(s); 1345 if (use != NULL) { 1346 return data_size(use); 1347 } 1348 } 1349 1350 int bsize = type2aelembytes(velt_basic_type(s)); 1351 assert(bsize != 0, "valid size"); 1352 return bsize; 1353 } 1354 1355 //------------------------------extend_packlist--------------------------- 1356 // Extend packset by following use->def and def->use links from pack members. 1357 void SuperWord::extend_packlist() { 1358 bool changed; 1359 do { 1360 packset_sort(_packset.length()); 1361 changed = false; 1362 for (int i = 0; i < _packset.length(); i++) { 1363 Node_List* p = _packset.at(i); 1364 changed |= follow_use_defs(p); 1365 changed |= follow_def_uses(p); 1366 } 1367 } while (changed); 1368 1369 if (_race_possible) { 1370 for (int i = 0; i < _packset.length(); i++) { 1371 Node_List* p = _packset.at(i); 1372 order_def_uses(p); 1373 } 1374 } 1375 1376 if (TraceSuperWord) { 1377 tty->print_cr("\nAfter extend_packlist"); 1378 print_packset(); 1379 } 1380 } 1381 1382 //------------------------------follow_use_defs--------------------------- 1383 // Extend the packset by visiting operand definitions of nodes in pack p 1384 bool SuperWord::follow_use_defs(Node_List* p) { 1385 assert(p->size() == 2, "just checking"); 1386 Node* s1 = p->at(0); 1387 Node* s2 = p->at(1); 1388 assert(s1->req() == s2->req(), "just checking"); 1389 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking"); 1390 1391 if (s1->is_Load()) return false; 1392 1393 int align = alignment(s1); 1394 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);) 1395 bool changed = false; 1396 int start = s1->is_Store() ? MemNode::ValueIn : 1; 1397 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req(); 1398 for (int j = start; j < end; j++) { 1399 Node* t1 = s1->in(j); 1400 Node* t2 = s2->in(j); 1401 if (!in_bb(t1) || !in_bb(t2)) 1402 continue; 1403 if (stmts_can_pack(t1, t2, align)) { 1404 if (est_savings(t1, t2) >= 0) { 1405 Node_List* pair = new Node_List(); 1406 pair->push(t1); 1407 pair->push(t2); 1408 _packset.append(pair); 1409 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);) 1410 set_alignment(t1, t2, align); 1411 changed = true; 1412 } 1413 } 1414 } 1415 return changed; 1416 } 1417 1418 //------------------------------follow_def_uses--------------------------- 1419 // Extend the packset by visiting uses of nodes in pack p 1420 bool SuperWord::follow_def_uses(Node_List* p) { 1421 bool changed = false; 1422 Node* s1 = p->at(0); 1423 Node* s2 = p->at(1); 1424 assert(p->size() == 2, "just checking"); 1425 assert(s1->req() == s2->req(), "just checking"); 1426 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking"); 1427 1428 if (s1->is_Store()) return false; 1429 1430 int align = alignment(s1); 1431 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);) 1432 int savings = -1; 1433 int num_s1_uses = 0; 1434 Node* u1 = NULL; 1435 Node* u2 = NULL; 1436 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { 1437 Node* t1 = s1->fast_out(i); 1438 num_s1_uses++; 1439 if (!in_bb(t1)) continue; 1440 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) { 1441 Node* t2 = s2->fast_out(j); 1442 if (!in_bb(t2)) continue; 1443 if (t2->Opcode() == Op_AddI && t2 == _lp->as_CountedLoop()->incr()) continue; // don't mess with the iv 1444 if (!opnd_positions_match(s1, t1, s2, t2)) 1445 continue; 1446 if (stmts_can_pack(t1, t2, align)) { 1447 int my_savings = est_savings(t1, t2); 1448 if (my_savings > savings) { 1449 savings = my_savings; 1450 u1 = t1; 1451 u2 = t2; 1452 } 1453 } 1454 } 1455 } 1456 if (num_s1_uses > 1) { 1457 _race_possible = true; 1458 } 1459 if (savings >= 0) { 1460 Node_List* pair = new Node_List(); 1461 pair->push(u1); 1462 pair->push(u2); 1463 _packset.append(pair); 1464 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);) 1465 set_alignment(u1, u2, align); 1466 changed = true; 1467 } 1468 return changed; 1469 } 1470 1471 //------------------------------order_def_uses--------------------------- 1472 // For extended packsets, ordinally arrange uses packset by major component 1473 void SuperWord::order_def_uses(Node_List* p) { 1474 Node* s1 = p->at(0); 1475 1476 if (s1->is_Store()) return; 1477 1478 // reductions are always managed beforehand 1479 if (s1->is_reduction()) return; 1480 1481 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { 1482 Node* t1 = s1->fast_out(i); 1483 1484 // Only allow operand swap on commuting operations 1485 if (!t1->is_Add() && !t1->is_Mul()) { 1486 break; 1487 } 1488 1489 // Now find t1's packset 1490 Node_List* p2 = NULL; 1491 for (int j = 0; j < _packset.length(); j++) { 1492 p2 = _packset.at(j); 1493 Node* first = p2->at(0); 1494 if (t1 == first) { 1495 break; 1496 } 1497 p2 = NULL; 1498 } 1499 // Arrange all sub components by the major component 1500 if (p2 != NULL) { 1501 for (uint j = 1; j < p->size(); j++) { 1502 Node* d1 = p->at(j); 1503 Node* u1 = p2->at(j); 1504 opnd_positions_match(s1, t1, d1, u1); 1505 } 1506 } 1507 } 1508 } 1509 1510 //---------------------------opnd_positions_match------------------------- 1511 // Is the use of d1 in u1 at the same operand position as d2 in u2? 1512 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) { 1513 // check reductions to see if they are marshalled to represent the reduction 1514 // operator in a specified opnd 1515 if (u1->is_reduction() && u2->is_reduction()) { 1516 // ensure reductions have phis and reduction definitions feeding the 1st operand 1517 Node* first = u1->in(2); 1518 if (first->is_Phi() || first->is_reduction()) { 1519 u1->swap_edges(1, 2); 1520 } 1521 // ensure reductions have phis and reduction definitions feeding the 1st operand 1522 first = u2->in(2); 1523 if (first->is_Phi() || first->is_reduction()) { 1524 u2->swap_edges(1, 2); 1525 } 1526 return true; 1527 } 1528 1529 uint ct = u1->req(); 1530 if (ct != u2->req()) return false; 1531 uint i1 = 0; 1532 uint i2 = 0; 1533 do { 1534 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break; 1535 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break; 1536 if (i1 != i2) { 1537 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) { 1538 // Further analysis relies on operands position matching. 1539 u2->swap_edges(i1, i2); 1540 } else { 1541 return false; 1542 } 1543 } 1544 } while (i1 < ct); 1545 return true; 1546 } 1547 1548 //------------------------------est_savings--------------------------- 1549 // Estimate the savings from executing s1 and s2 as a pack 1550 int SuperWord::est_savings(Node* s1, Node* s2) { 1551 int save_in = 2 - 1; // 2 operations per instruction in packed form 1552 1553 // inputs 1554 for (uint i = 1; i < s1->req(); i++) { 1555 Node* x1 = s1->in(i); 1556 Node* x2 = s2->in(i); 1557 if (x1 != x2) { 1558 if (are_adjacent_refs(x1, x2)) { 1559 save_in += adjacent_profit(x1, x2); 1560 } else if (!in_packset(x1, x2)) { 1561 save_in -= pack_cost(2); 1562 } else { 1563 save_in += unpack_cost(2); 1564 } 1565 } 1566 } 1567 1568 // uses of result 1569 uint ct = 0; 1570 int save_use = 0; 1571 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) { 1572 Node* s1_use = s1->fast_out(i); 1573 for (int j = 0; j < _packset.length(); j++) { 1574 Node_List* p = _packset.at(j); 1575 if (p->at(0) == s1_use) { 1576 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) { 1577 Node* s2_use = s2->fast_out(k); 1578 if (p->at(p->size()-1) == s2_use) { 1579 ct++; 1580 if (are_adjacent_refs(s1_use, s2_use)) { 1581 save_use += adjacent_profit(s1_use, s2_use); 1582 } 1583 } 1584 } 1585 } 1586 } 1587 } 1588 1589 if (ct < s1->outcnt()) save_use += unpack_cost(1); 1590 if (ct < s2->outcnt()) save_use += unpack_cost(1); 1591 1592 return MAX2(save_in, save_use); 1593 } 1594 1595 //------------------------------costs--------------------------- 1596 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; } 1597 int SuperWord::pack_cost(int ct) { return ct; } 1598 int SuperWord::unpack_cost(int ct) { return ct; } 1599 1600 //------------------------------combine_packs--------------------------- 1601 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last 1602 void SuperWord::combine_packs() { 1603 bool changed = true; 1604 // Combine packs regardless max vector size. 1605 while (changed) { 1606 changed = false; 1607 for (int i = 0; i < _packset.length(); i++) { 1608 Node_List* p1 = _packset.at(i); 1609 if (p1 == NULL) continue; 1610 // Because of sorting we can start at i + 1 1611 for (int j = i + 1; j < _packset.length(); j++) { 1612 Node_List* p2 = _packset.at(j); 1613 if (p2 == NULL) continue; 1614 if (i == j) continue; 1615 if (p1->at(p1->size()-1) == p2->at(0)) { 1616 for (uint k = 1; k < p2->size(); k++) { 1617 p1->push(p2->at(k)); 1618 } 1619 _packset.at_put(j, NULL); 1620 changed = true; 1621 } 1622 } 1623 } 1624 } 1625 1626 // Split packs which have size greater then max vector size. 1627 for (int i = 0; i < _packset.length(); i++) { 1628 Node_List* p1 = _packset.at(i); 1629 if (p1 != NULL) { 1630 BasicType bt = velt_basic_type(p1->at(0)); 1631 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector 1632 assert(is_power_of_2(max_vlen), "sanity"); 1633 uint psize = p1->size(); 1634 if (!is_power_of_2(psize)) { 1635 // Skip pack which can't be vector. 1636 // case1: for(...) { a[i] = i; } elements values are different (i+x) 1637 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store 1638 _packset.at_put(i, NULL); 1639 continue; 1640 } 1641 if (psize > max_vlen) { 1642 Node_List* pack = new Node_List(); 1643 for (uint j = 0; j < psize; j++) { 1644 pack->push(p1->at(j)); 1645 if (pack->size() >= max_vlen) { 1646 assert(is_power_of_2(pack->size()), "sanity"); 1647 _packset.append(pack); 1648 pack = new Node_List(); 1649 } 1650 } 1651 _packset.at_put(i, NULL); 1652 } 1653 } 1654 } 1655 1656 // Compress list. 1657 for (int i = _packset.length() - 1; i >= 0; i--) { 1658 Node_List* p1 = _packset.at(i); 1659 if (p1 == NULL) { 1660 _packset.remove_at(i); 1661 } 1662 } 1663 1664 if (TraceSuperWord) { 1665 tty->print_cr("\nAfter combine_packs"); 1666 print_packset(); 1667 } 1668 } 1669 1670 //-----------------------------construct_my_pack_map-------------------------- 1671 // Construct the map from nodes to packs. Only valid after the 1672 // point where a node is only in one pack (after combine_packs). 1673 void SuperWord::construct_my_pack_map() { 1674 Node_List* rslt = NULL; 1675 for (int i = 0; i < _packset.length(); i++) { 1676 Node_List* p = _packset.at(i); 1677 for (uint j = 0; j < p->size(); j++) { 1678 Node* s = p->at(j); 1679 assert(my_pack(s) == NULL, "only in one pack"); 1680 set_my_pack(s, p); 1681 } 1682 } 1683 } 1684 1685 //------------------------------filter_packs--------------------------- 1686 // Remove packs that are not implemented or not profitable. 1687 void SuperWord::filter_packs() { 1688 // Remove packs that are not implemented 1689 for (int i = _packset.length() - 1; i >= 0; i--) { 1690 Node_List* pk = _packset.at(i); 1691 bool impl = implemented(pk); 1692 if (!impl) { 1693 #ifndef PRODUCT 1694 if (TraceSuperWord && Verbose) { 1695 tty->print_cr("Unimplemented"); 1696 pk->at(0)->dump(); 1697 } 1698 #endif 1699 remove_pack_at(i); 1700 } 1701 Node *n = pk->at(0); 1702 if (n->is_reduction()) { 1703 _num_reductions++; 1704 } else { 1705 _num_work_vecs++; 1706 } 1707 } 1708 1709 // Remove packs that are not profitable 1710 bool changed; 1711 do { 1712 changed = false; 1713 for (int i = _packset.length() - 1; i >= 0; i--) { 1714 Node_List* pk = _packset.at(i); 1715 bool prof = profitable(pk); 1716 if (!prof) { 1717 #ifndef PRODUCT 1718 if (TraceSuperWord && Verbose) { 1719 tty->print_cr("Unprofitable"); 1720 pk->at(0)->dump(); 1721 } 1722 #endif 1723 remove_pack_at(i); 1724 changed = true; 1725 } 1726 } 1727 } while (changed); 1728 1729 #ifndef PRODUCT 1730 if (TraceSuperWord) { 1731 tty->print_cr("\nAfter filter_packs"); 1732 print_packset(); 1733 tty->cr(); 1734 } 1735 #endif 1736 } 1737 1738 //------------------------------merge_packs_to_cmovd--------------------------- 1739 // Merge CMoveD into new vector-nodes 1740 // We want to catch this pattern and subsume CmpD and Bool into CMoveD 1741 // 1742 // SubD ConD 1743 // / | / 1744 // / | / / 1745 // / | / / 1746 // / | / / 1747 // / / / 1748 // / / | / 1749 // v / | / 1750 // CmpD | / 1751 // | | / 1752 // v | / 1753 // Bool | / 1754 // \ | / 1755 // \ | / 1756 // \ | / 1757 // \ | / 1758 // \ v / 1759 // CMoveD 1760 // 1761 1762 void SuperWord::merge_packs_to_cmovd() { 1763 for (int i = _packset.length() - 1; i >= 0; i--) { 1764 _cmovev_kit.make_cmovevd_pack(_packset.at(i)); 1765 } 1766 #ifndef PRODUCT 1767 if (TraceSuperWord) { 1768 tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge"); 1769 print_packset(); 1770 tty->cr(); 1771 } 1772 #endif 1773 } 1774 1775 Node* CMoveKit::is_Bool_candidate(Node* def) const { 1776 Node* use = NULL; 1777 if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) { 1778 return NULL; 1779 } 1780 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 1781 use = def->fast_out(j); 1782 if (!_sw->same_generation(def, use) || !use->is_CMove()) { 1783 return NULL; 1784 } 1785 } 1786 return use; 1787 } 1788 1789 Node* CMoveKit::is_CmpD_candidate(Node* def) const { 1790 Node* use = NULL; 1791 if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) { 1792 return NULL; 1793 } 1794 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 1795 use = def->fast_out(j); 1796 if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) { 1797 return NULL; 1798 } 1799 } 1800 return use; 1801 } 1802 1803 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) { 1804 Node *cmovd = cmovd_pk->at(0); 1805 if (!cmovd->is_CMove()) { 1806 return NULL; 1807 } 1808 if (cmovd->Opcode() != Op_CMoveF && cmovd->Opcode() != Op_CMoveD) { 1809 return NULL; 1810 } 1811 if (pack(cmovd) != NULL) { // already in the cmov pack 1812 return NULL; 1813 } 1814 if (cmovd->in(0) != NULL) { 1815 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();}) 1816 return NULL; 1817 } 1818 1819 Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition); 1820 if (!bol->is_Bool() 1821 || bol->outcnt() != 1 1822 || !_sw->same_generation(bol, cmovd) 1823 || bol->in(0) != NULL // BoolNode has control flow!! 1824 || _sw->my_pack(bol) == NULL) { 1825 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: Bool %d does not fit CMoveD %d for building vector, escaping...", bol->_idx, cmovd->_idx); bol->dump();}) 1826 return NULL; 1827 } 1828 Node_List* bool_pk = _sw->my_pack(bol); 1829 if (bool_pk->size() != cmovd_pk->size() ) { 1830 return NULL; 1831 } 1832 1833 Node* cmpd = bol->in(1); 1834 if (!cmpd->is_Cmp() 1835 || cmpd->outcnt() != 1 1836 || !_sw->same_generation(cmpd, cmovd) 1837 || cmpd->in(0) != NULL // CmpDNode has control flow!! 1838 || _sw->my_pack(cmpd) == NULL) { 1839 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CmpD %d does not fit CMoveD %d for building vector, escaping...", cmpd->_idx, cmovd->_idx); cmpd->dump();}) 1840 return NULL; 1841 } 1842 Node_List* cmpd_pk = _sw->my_pack(cmpd); 1843 if (cmpd_pk->size() != cmovd_pk->size() ) { 1844 return NULL; 1845 } 1846 1847 if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) { 1848 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: cmpd pack for CmpD %d failed vectorization test", cmpd->_idx); cmpd->dump();}) 1849 return NULL; 1850 } 1851 1852 Node_List* new_cmpd_pk = new Node_List(); 1853 uint sz = cmovd_pk->size() - 1; 1854 for (uint i = 0; i <= sz; ++i) { 1855 Node* cmov = cmovd_pk->at(i); 1856 Node* bol = bool_pk->at(i); 1857 Node* cmp = cmpd_pk->at(i); 1858 1859 new_cmpd_pk->insert(i, cmov); 1860 1861 map(cmov, new_cmpd_pk); 1862 map(bol, new_cmpd_pk); 1863 map(cmp, new_cmpd_pk); 1864 1865 _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool 1866 } 1867 _sw->_packset.remove(cmovd_pk); 1868 _sw->_packset.remove(bool_pk); 1869 _sw->_packset.remove(cmpd_pk); 1870 _sw->_packset.append(new_cmpd_pk); 1871 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);}) 1872 return new_cmpd_pk; 1873 } 1874 1875 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) { 1876 Node* cmpd0 = cmpd_pk->at(0); 1877 assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode"); 1878 assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD"); 1879 assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size"); 1880 Node* in1 = cmpd0->in(1); 1881 Node* in2 = cmpd0->in(2); 1882 Node_List* in1_pk = _sw->my_pack(in1); 1883 Node_List* in2_pk = _sw->my_pack(in2); 1884 1885 if ( (in1_pk != NULL && in1_pk->size() != cmpd_pk->size()) 1886 || (in2_pk != NULL && in2_pk->size() != cmpd_pk->size()) ) { 1887 return false; 1888 } 1889 1890 // test if "all" in1 are in the same pack or the same node 1891 if (in1_pk == NULL) { 1892 for (uint j = 1; j < cmpd_pk->size(); j++) { 1893 if (cmpd_pk->at(j)->in(1) != in1) { 1894 return false; 1895 } 1896 }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1) 1897 } 1898 // test if "all" in2 are in the same pack or the same node 1899 if (in2_pk == NULL) { 1900 for (uint j = 1; j < cmpd_pk->size(); j++) { 1901 if (cmpd_pk->at(j)->in(2) != in2) { 1902 return false; 1903 } 1904 }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2) 1905 } 1906 //now check if cmpd_pk may be subsumed in vector built for cmovd_pk 1907 int cmovd_ind1, cmovd_ind2; 1908 if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse) 1909 && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) { 1910 cmovd_ind1 = CMoveNode::IfFalse; 1911 cmovd_ind2 = CMoveNode::IfTrue; 1912 } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse) 1913 && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) { 1914 cmovd_ind2 = CMoveNode::IfFalse; 1915 cmovd_ind1 = CMoveNode::IfTrue; 1916 } 1917 else { 1918 return false; 1919 } 1920 1921 for (uint j = 1; j < cmpd_pk->size(); j++) { 1922 if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1) 1923 || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) { 1924 return false; 1925 }//if 1926 } 1927 NOT_PRODUCT(if(_sw->is_trace_cmov()) { tty->print("CMoveKit::test_cmpd_pack: cmpd pack for 1st CmpD %d is OK for vectorization: ", cmpd0->_idx); cmpd0->dump(); }) 1928 return true; 1929 } 1930 1931 //------------------------------implemented--------------------------- 1932 // Can code be generated for pack p? 1933 bool SuperWord::implemented(Node_List* p) { 1934 bool retValue = false; 1935 Node* p0 = p->at(0); 1936 if (p0 != NULL) { 1937 int opc = p0->Opcode(); 1938 uint size = p->size(); 1939 if (p0->is_reduction()) { 1940 const Type *arith_type = p0->bottom_type(); 1941 // Length 2 reductions of INT/LONG do not offer performance benefits 1942 if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) { 1943 retValue = false; 1944 } else { 1945 retValue = ReductionNode::implemented(opc, size, arith_type->basic_type()); 1946 } 1947 } else { 1948 retValue = VectorNode::implemented(opc, size, velt_basic_type(p0)); 1949 } 1950 if (!retValue) { 1951 if (is_cmov_pack(p)) { 1952 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);}) 1953 return true; 1954 } 1955 } 1956 } 1957 return retValue; 1958 } 1959 1960 bool SuperWord::is_cmov_pack(Node_List* p) { 1961 return _cmovev_kit.pack(p->at(0)) != NULL; 1962 } 1963 //------------------------------same_inputs-------------------------- 1964 // For pack p, are all idx operands the same? 1965 bool SuperWord::same_inputs(Node_List* p, int idx) { 1966 Node* p0 = p->at(0); 1967 uint vlen = p->size(); 1968 Node* p0_def = p0->in(idx); 1969 for (uint i = 1; i < vlen; i++) { 1970 Node* pi = p->at(i); 1971 Node* pi_def = pi->in(idx); 1972 if (p0_def != pi_def) { 1973 return false; 1974 } 1975 } 1976 return true; 1977 } 1978 1979 //------------------------------profitable--------------------------- 1980 // For pack p, are all operands and all uses (with in the block) vector? 1981 bool SuperWord::profitable(Node_List* p) { 1982 Node* p0 = p->at(0); 1983 uint start, end; 1984 VectorNode::vector_operands(p0, &start, &end); 1985 1986 // Return false if some inputs are not vectors or vectors with different 1987 // size or alignment. 1988 // Also, for now, return false if not scalar promotion case when inputs are 1989 // the same. Later, implement PackNode and allow differing, non-vector inputs 1990 // (maybe just the ones from outside the block.) 1991 for (uint i = start; i < end; i++) { 1992 if (!is_vector_use(p0, i)) { 1993 return false; 1994 } 1995 } 1996 // Check if reductions are connected 1997 if (p0->is_reduction()) { 1998 Node* second_in = p0->in(2); 1999 Node_List* second_pk = my_pack(second_in); 2000 if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) { 2001 // Remove reduction flag if no parent pack or if not enough work 2002 // to cover reduction expansion overhead 2003 p0->remove_flag(Node::Flag_is_reduction); 2004 return false; 2005 } else if (second_pk->size() != p->size()) { 2006 return false; 2007 } 2008 } 2009 if (VectorNode::is_shift(p0)) { 2010 // For now, return false if shift count is vector or not scalar promotion 2011 // case (different shift counts) because it is not supported yet. 2012 Node* cnt = p0->in(2); 2013 Node_List* cnt_pk = my_pack(cnt); 2014 if (cnt_pk != NULL) 2015 return false; 2016 if (!same_inputs(p, 2)) 2017 return false; 2018 } 2019 if (!p0->is_Store()) { 2020 // For now, return false if not all uses are vector. 2021 // Later, implement ExtractNode and allow non-vector uses (maybe 2022 // just the ones outside the block.) 2023 for (uint i = 0; i < p->size(); i++) { 2024 Node* def = p->at(i); 2025 if (is_cmov_pack_internal_node(p, def)) { 2026 continue; 2027 } 2028 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2029 Node* use = def->fast_out(j); 2030 for (uint k = 0; k < use->req(); k++) { 2031 Node* n = use->in(k); 2032 if (def == n) { 2033 // reductions should only have a Phi use at the the loop 2034 // head and out of loop uses 2035 if (def->is_reduction() && 2036 ((use->is_Phi() && use->in(0) == _lpt->_head) || 2037 !_lpt->is_member(_phase->get_loop(_phase->ctrl_or_self(use))))) { 2038 assert(i == p->size()-1, "must be last element of the pack"); 2039 continue; 2040 } 2041 if (!is_vector_use(use, k)) { 2042 return false; 2043 } 2044 } 2045 } 2046 } 2047 } 2048 } 2049 return true; 2050 } 2051 2052 //------------------------------schedule--------------------------- 2053 // Adjust the memory graph for the packed operations 2054 void SuperWord::schedule() { 2055 2056 // Co-locate in the memory graph the members of each memory pack 2057 for (int i = 0; i < _packset.length(); i++) { 2058 co_locate_pack(_packset.at(i)); 2059 } 2060 } 2061 2062 //-------------------------------remove_and_insert------------------- 2063 // Remove "current" from its current position in the memory graph and insert 2064 // it after the appropriate insertion point (lip or uip). 2065 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip, 2066 Node *uip, Unique_Node_List &sched_before) { 2067 Node* my_mem = current->in(MemNode::Memory); 2068 bool sched_up = sched_before.member(current); 2069 2070 // remove current_store from its current position in the memmory graph 2071 for (DUIterator i = current->outs(); current->has_out(i); i++) { 2072 Node* use = current->out(i); 2073 if (use->is_Mem()) { 2074 assert(use->in(MemNode::Memory) == current, "must be"); 2075 if (use == prev) { // connect prev to my_mem 2076 _igvn.replace_input_of(use, MemNode::Memory, my_mem); 2077 --i; //deleted this edge; rescan position 2078 } else if (sched_before.member(use)) { 2079 if (!sched_up) { // Will be moved together with current 2080 _igvn.replace_input_of(use, MemNode::Memory, uip); 2081 --i; //deleted this edge; rescan position 2082 } 2083 } else { 2084 if (sched_up) { // Will be moved together with current 2085 _igvn.replace_input_of(use, MemNode::Memory, lip); 2086 --i; //deleted this edge; rescan position 2087 } 2088 } 2089 } 2090 } 2091 2092 Node *insert_pt = sched_up ? uip : lip; 2093 2094 // all uses of insert_pt's memory state should use current's instead 2095 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) { 2096 Node* use = insert_pt->out(i); 2097 if (use->is_Mem()) { 2098 assert(use->in(MemNode::Memory) == insert_pt, "must be"); 2099 _igvn.replace_input_of(use, MemNode::Memory, current); 2100 --i; //deleted this edge; rescan position 2101 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) { 2102 uint pos; //lip (lower insert point) must be the last one in the memory slice 2103 for (pos=1; pos < use->req(); pos++) { 2104 if (use->in(pos) == insert_pt) break; 2105 } 2106 _igvn.replace_input_of(use, pos, current); 2107 --i; 2108 } 2109 } 2110 2111 //connect current to insert_pt 2112 _igvn.replace_input_of(current, MemNode::Memory, insert_pt); 2113 } 2114 2115 //------------------------------co_locate_pack---------------------------------- 2116 // To schedule a store pack, we need to move any sandwiched memory ops either before 2117 // or after the pack, based upon dependence information: 2118 // (1) If any store in the pack depends on the sandwiched memory op, the 2119 // sandwiched memory op must be scheduled BEFORE the pack; 2120 // (2) If a sandwiched memory op depends on any store in the pack, the 2121 // sandwiched memory op must be scheduled AFTER the pack; 2122 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched 2123 // memory op (say memB), memB must be scheduled before memA. So, if memA is 2124 // scheduled before the pack, memB must also be scheduled before the pack; 2125 // (4) If there is no dependence restriction for a sandwiched memory op, we simply 2126 // schedule this store AFTER the pack 2127 // (5) We know there is no dependence cycle, so there in no other case; 2128 // (6) Finally, all memory ops in another single pack should be moved in the same direction. 2129 // 2130 // To schedule a load pack, we use the memory state of either the first or the last load in 2131 // the pack, based on the dependence constraint. 2132 void SuperWord::co_locate_pack(Node_List* pk) { 2133 if (pk->at(0)->is_Store()) { 2134 MemNode* first = executed_first(pk)->as_Mem(); 2135 MemNode* last = executed_last(pk)->as_Mem(); 2136 Unique_Node_List schedule_before_pack; 2137 Unique_Node_List memops; 2138 2139 MemNode* current = last->in(MemNode::Memory)->as_Mem(); 2140 MemNode* previous = last; 2141 while (true) { 2142 assert(in_bb(current), "stay in block"); 2143 memops.push(previous); 2144 for (DUIterator i = current->outs(); current->has_out(i); i++) { 2145 Node* use = current->out(i); 2146 if (use->is_Mem() && use != previous) 2147 memops.push(use); 2148 } 2149 if (current == first) break; 2150 previous = current; 2151 current = current->in(MemNode::Memory)->as_Mem(); 2152 } 2153 2154 // determine which memory operations should be scheduled before the pack 2155 for (uint i = 1; i < memops.size(); i++) { 2156 Node *s1 = memops.at(i); 2157 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) { 2158 for (uint j = 0; j< i; j++) { 2159 Node *s2 = memops.at(j); 2160 if (!independent(s1, s2)) { 2161 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) { 2162 schedule_before_pack.push(s1); // s1 must be scheduled before 2163 Node_List* mem_pk = my_pack(s1); 2164 if (mem_pk != NULL) { 2165 for (uint ii = 0; ii < mem_pk->size(); ii++) { 2166 Node* s = mem_pk->at(ii); // follow partner 2167 if (memops.member(s) && !schedule_before_pack.member(s)) 2168 schedule_before_pack.push(s); 2169 } 2170 } 2171 break; 2172 } 2173 } 2174 } 2175 } 2176 } 2177 2178 Node* upper_insert_pt = first->in(MemNode::Memory); 2179 // Following code moves loads connected to upper_insert_pt below aliased stores. 2180 // Collect such loads here and reconnect them back to upper_insert_pt later. 2181 memops.clear(); 2182 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) { 2183 Node* use = upper_insert_pt->out(i); 2184 if (use->is_Mem() && !use->is_Store()) { 2185 memops.push(use); 2186 } 2187 } 2188 2189 MemNode* lower_insert_pt = last; 2190 previous = last; //previous store in pk 2191 current = last->in(MemNode::Memory)->as_Mem(); 2192 2193 // start scheduling from "last" to "first" 2194 while (true) { 2195 assert(in_bb(current), "stay in block"); 2196 assert(in_pack(previous, pk), "previous stays in pack"); 2197 Node* my_mem = current->in(MemNode::Memory); 2198 2199 if (in_pack(current, pk)) { 2200 // Forward users of my memory state (except "previous) to my input memory state 2201 for (DUIterator i = current->outs(); current->has_out(i); i++) { 2202 Node* use = current->out(i); 2203 if (use->is_Mem() && use != previous) { 2204 assert(use->in(MemNode::Memory) == current, "must be"); 2205 if (schedule_before_pack.member(use)) { 2206 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt); 2207 } else { 2208 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt); 2209 } 2210 --i; // deleted this edge; rescan position 2211 } 2212 } 2213 previous = current; 2214 } else { // !in_pack(current, pk) ==> a sandwiched store 2215 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack); 2216 } 2217 2218 if (current == first) break; 2219 current = my_mem->as_Mem(); 2220 } // end while 2221 2222 // Reconnect loads back to upper_insert_pt. 2223 for (uint i = 0; i < memops.size(); i++) { 2224 Node *ld = memops.at(i); 2225 if (ld->in(MemNode::Memory) != upper_insert_pt) { 2226 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt); 2227 } 2228 } 2229 } else if (pk->at(0)->is_Load()) { //load 2230 // all loads in the pack should have the same memory state. By default, 2231 // we use the memory state of the last load. However, if any load could 2232 // not be moved down due to the dependence constraint, we use the memory 2233 // state of the first load. 2234 Node* first_mem = pk->at(0)->in(MemNode::Memory); 2235 Node* last_mem = first_mem; 2236 for (uint i = 1; i < pk->size(); i++) { 2237 Node* ld = pk->at(i); 2238 Node* mem = ld->in(MemNode::Memory); 2239 assert(in_bb(first_mem) || in_bb(mem) || mem == first_mem, "2 different memory state from outside the loop?"); 2240 if (in_bb(mem)) { 2241 if (in_bb(first_mem) && bb_idx(mem) < bb_idx(first_mem)) { 2242 first_mem = mem; 2243 } 2244 if (!in_bb(last_mem) || bb_idx(mem) > bb_idx(last_mem)) { 2245 last_mem = mem; 2246 } 2247 } 2248 } 2249 bool schedule_last = true; 2250 for (uint i = 0; i < pk->size(); i++) { 2251 Node* ld = pk->at(i); 2252 for (Node* current = last_mem; current != ld->in(MemNode::Memory); 2253 current=current->in(MemNode::Memory)) { 2254 assert(current != first_mem, "corrupted memory graph"); 2255 if(current->is_Mem() && !independent(current, ld)){ 2256 schedule_last = false; // a later store depends on this load 2257 break; 2258 } 2259 } 2260 } 2261 2262 Node* mem_input = schedule_last ? last_mem : first_mem; 2263 _igvn.hash_delete(mem_input); 2264 // Give each load the same memory state 2265 for (uint i = 0; i < pk->size(); i++) { 2266 LoadNode* ld = pk->at(i)->as_Load(); 2267 _igvn.replace_input_of(ld, MemNode::Memory, mem_input); 2268 } 2269 } 2270 } 2271 2272 #ifndef PRODUCT 2273 void SuperWord::print_loop(bool whole) { 2274 Node_Stack stack(_arena, _phase->C->unique() >> 2); 2275 Node_List rpo_list; 2276 VectorSet visited(_arena); 2277 visited.set(lpt()->_head->_idx); 2278 _phase->rpo(lpt()->_head, stack, visited, rpo_list); 2279 _phase->dump(lpt(), rpo_list.size(), rpo_list ); 2280 if(whole) { 2281 tty->print_cr("\n Whole loop tree"); 2282 _phase->dump(); 2283 tty->print_cr(" End of whole loop tree\n"); 2284 } 2285 } 2286 #endif 2287 2288 //------------------------------output--------------------------- 2289 // Convert packs into vector node operations 2290 void SuperWord::output() { 2291 CountedLoopNode *cl = lpt()->_head->as_CountedLoop(); 2292 Compile* C = _phase->C; 2293 if (_packset.length() == 0) { 2294 if (cl->is_main_loop()) { 2295 // Instigate more unrolling for optimization when vectorization fails. 2296 C->set_major_progress(); 2297 cl->set_notpassed_slp(); 2298 cl->mark_do_unroll_only(); 2299 } 2300 return; 2301 } 2302 2303 #ifndef PRODUCT 2304 if (TraceLoopOpts) { 2305 tty->print("SuperWord::output "); 2306 lpt()->dump_head(); 2307 } 2308 #endif 2309 2310 if (cl->is_main_loop()) { 2311 // MUST ENSURE main loop's initial value is properly aligned: 2312 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0 2313 2314 align_initial_loop_index(align_to_ref()); 2315 2316 // Insert extract (unpack) operations for scalar uses 2317 for (int i = 0; i < _packset.length(); i++) { 2318 insert_extracts(_packset.at(i)); 2319 } 2320 } 2321 2322 uint max_vlen_in_bytes = 0; 2323 uint max_vlen = 0; 2324 bool can_process_post_loop = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()); 2325 2326 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);}) 2327 2328 CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy()); 2329 2330 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);}) 2331 2332 if (do_reserve_copy() && !make_reversable.has_reserved()) { 2333 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");}) 2334 return; 2335 } 2336 2337 for (int i = 0; i < _block.length(); i++) { 2338 Node* n = _block.at(i); 2339 Node_List* p = my_pack(n); 2340 if (p && n == executed_last(p)) { 2341 uint vlen = p->size(); 2342 uint vlen_in_bytes = 0; 2343 Node* vn = NULL; 2344 Node* low_adr = p->at(0); 2345 Node* first = executed_first(p); 2346 if (can_process_post_loop) { 2347 // override vlen with the main loops vector length 2348 vlen = cl->slp_max_unroll(); 2349 } 2350 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: %d executed first, %d executed last in pack", first->_idx, n->_idx); print_pack(p);}) 2351 int opc = n->Opcode(); 2352 if (n->is_Load()) { 2353 Node* ctl = n->in(MemNode::Control); 2354 Node* mem = first->in(MemNode::Memory); 2355 SWPointer p1(n->as_Mem(), this, NULL, false); 2356 // Identify the memory dependency for the new loadVector node by 2357 // walking up through memory chain. 2358 // This is done to give flexibility to the new loadVector node so that 2359 // it can move above independent storeVector nodes. 2360 while (mem->is_StoreVector()) { 2361 SWPointer p2(mem->as_Mem(), this, NULL, false); 2362 int cmp = p1.cmp(p2); 2363 if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) { 2364 mem = mem->in(MemNode::Memory); 2365 } else { 2366 break; // dependent memory 2367 } 2368 } 2369 Node* adr = low_adr->in(MemNode::Address); 2370 const TypePtr* atyp = n->adr_type(); 2371 vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p)); 2372 vlen_in_bytes = vn->as_LoadVector()->memory_size(); 2373 } else if (n->is_Store()) { 2374 // Promote value to be stored to vector 2375 Node* val = vector_opd(p, MemNode::ValueIn); 2376 if (val == NULL) { 2377 if (do_reserve_copy()) { 2378 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");}) 2379 return; //and reverse to backup IG 2380 } 2381 ShouldNotReachHere(); 2382 } 2383 2384 Node* ctl = n->in(MemNode::Control); 2385 Node* mem = first->in(MemNode::Memory); 2386 Node* adr = low_adr->in(MemNode::Address); 2387 const TypePtr* atyp = n->adr_type(); 2388 vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen); 2389 vlen_in_bytes = vn->as_StoreVector()->memory_size(); 2390 } else if (VectorNode::is_muladds2i(n)) { 2391 assert(n->req() == 5u, "MulAddS2I should have 4 operands."); 2392 Node* in1 = vector_opd(p, 1); 2393 Node* in2 = vector_opd(p, 2); 2394 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n)); 2395 vlen_in_bytes = vn->as_Vector()->length_in_bytes(); 2396 } else if (n->req() == 3 && !is_cmov_pack(p)) { 2397 // Promote operands to vector 2398 Node* in1 = NULL; 2399 bool node_isa_reduction = n->is_reduction(); 2400 if (node_isa_reduction) { 2401 // the input to the first reduction operation is retained 2402 in1 = low_adr->in(1); 2403 } else { 2404 in1 = vector_opd(p, 1); 2405 if (in1 == NULL) { 2406 if (do_reserve_copy()) { 2407 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");}) 2408 return; //and reverse to backup IG 2409 } 2410 ShouldNotReachHere(); 2411 } 2412 } 2413 Node* in2 = vector_opd(p, 2); 2414 if (in2 == NULL) { 2415 if (do_reserve_copy()) { 2416 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");}) 2417 return; //and reverse to backup IG 2418 } 2419 ShouldNotReachHere(); 2420 } 2421 if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) { 2422 // Move invariant vector input into second position to avoid register spilling. 2423 Node* tmp = in1; 2424 in1 = in2; 2425 in2 = tmp; 2426 } 2427 if (node_isa_reduction) { 2428 const Type *arith_type = n->bottom_type(); 2429 vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type()); 2430 if (in2->is_Load()) { 2431 vlen_in_bytes = in2->as_LoadVector()->memory_size(); 2432 } else { 2433 vlen_in_bytes = in2->as_Vector()->length_in_bytes(); 2434 } 2435 } else { 2436 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n)); 2437 vlen_in_bytes = vn->as_Vector()->length_in_bytes(); 2438 } 2439 } else if (opc == Op_SqrtF || opc == Op_SqrtD || 2440 opc == Op_AbsF || opc == Op_AbsD || 2441 opc == Op_NegF || opc == Op_NegD || 2442 opc == Op_PopCountI) { 2443 assert(n->req() == 2, "only one input expected"); 2444 Node* in = vector_opd(p, 1); 2445 vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n)); 2446 vlen_in_bytes = vn->as_Vector()->length_in_bytes(); 2447 } else if (is_cmov_pack(p)) { 2448 if (can_process_post_loop) { 2449 // do not refactor of flow in post loop context 2450 return; 2451 } 2452 if (!n->is_CMove()) { 2453 continue; 2454 } 2455 // place here CMoveVDNode 2456 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);}) 2457 Node* bol = n->in(CMoveNode::Condition); 2458 if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) { 2459 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: %d is not Bool node, trying its in(1) node %d", bol->_idx, bol->in(1)->_idx); bol->dump(); bol->in(1)->dump();}) 2460 bol = bol->in(1); //may be ExtractNode 2461 } 2462 2463 assert(bol->is_Bool(), "should be BoolNode - too late to bail out!"); 2464 if (!bol->is_Bool()) { 2465 if (do_reserve_copy()) { 2466 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();}) 2467 return; //and reverse to backup IG 2468 } 2469 ShouldNotReachHere(); 2470 } 2471 2472 int cond = (int)bol->as_Bool()->_test._test; 2473 Node* in_cc = _igvn.intcon(cond); 2474 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();}) 2475 Node* cc = bol->clone(); 2476 cc->set_req(1, in_cc); 2477 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();}) 2478 2479 Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse 2480 if (src1 == NULL) { 2481 if (do_reserve_copy()) { 2482 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");}) 2483 return; //and reverse to backup IG 2484 } 2485 ShouldNotReachHere(); 2486 } 2487 Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue 2488 if (src2 == NULL) { 2489 if (do_reserve_copy()) { 2490 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");}) 2491 return; //and reverse to backup IG 2492 } 2493 ShouldNotReachHere(); 2494 } 2495 BasicType bt = velt_basic_type(n); 2496 const TypeVect* vt = TypeVect::make(bt, vlen); 2497 assert(bt == T_FLOAT || bt == T_DOUBLE, "Only vectorization for FP cmovs is supported"); 2498 if (bt == T_FLOAT) { 2499 vn = new CMoveVFNode(cc, src1, src2, vt); 2500 } else { 2501 assert(bt == T_DOUBLE, "Expected double"); 2502 vn = new CMoveVDNode(cc, src1, src2, vt); 2503 } 2504 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();}) 2505 } else if (opc == Op_FmaD || opc == Op_FmaF) { 2506 // Promote operands to vector 2507 Node* in1 = vector_opd(p, 1); 2508 Node* in2 = vector_opd(p, 2); 2509 Node* in3 = vector_opd(p, 3); 2510 vn = VectorNode::make(opc, in1, in2, in3, vlen, velt_basic_type(n)); 2511 vlen_in_bytes = vn->as_Vector()->length_in_bytes(); 2512 } else { 2513 if (do_reserve_copy()) { 2514 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");}) 2515 return; //and reverse to backup IG 2516 } 2517 ShouldNotReachHere(); 2518 } 2519 2520 assert(vn != NULL, "sanity"); 2521 if (vn == NULL) { 2522 if (do_reserve_copy()){ 2523 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");}) 2524 return; //and reverse to backup IG 2525 } 2526 ShouldNotReachHere(); 2527 } 2528 2529 _block.at_put(i, vn); 2530 _igvn.register_new_node_with_optimizer(vn); 2531 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0))); 2532 for (uint j = 0; j < p->size(); j++) { 2533 Node* pm = p->at(j); 2534 _igvn.replace_node(pm, vn); 2535 } 2536 _igvn._worklist.push(vn); 2537 2538 if (can_process_post_loop) { 2539 // first check if the vector size if the maximum vector which we can use on the machine, 2540 // other vector size have reduced values for predicated data mapping. 2541 if (vlen_in_bytes != (uint)MaxVectorSize) { 2542 return; 2543 } 2544 } 2545 2546 if (vlen_in_bytes >= max_vlen_in_bytes && vlen > max_vlen) { 2547 max_vlen = vlen; 2548 max_vlen_in_bytes = vlen_in_bytes; 2549 } 2550 #ifdef ASSERT 2551 if (TraceNewVectors) { 2552 tty->print("new Vector node: "); 2553 vn->dump(); 2554 } 2555 #endif 2556 } 2557 }//for (int i = 0; i < _block.length(); i++) 2558 2559 if (max_vlen_in_bytes > C->max_vector_size()) { 2560 C->set_max_vector_size(max_vlen_in_bytes); 2561 } 2562 if (max_vlen_in_bytes > 0) { 2563 cl->mark_loop_vectorized(); 2564 } 2565 2566 if (SuperWordLoopUnrollAnalysis) { 2567 if (cl->has_passed_slp()) { 2568 uint slp_max_unroll_factor = cl->slp_max_unroll(); 2569 if (slp_max_unroll_factor == max_vlen) { 2570 if (TraceSuperWordLoopUnrollAnalysis) { 2571 tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte); 2572 } 2573 2574 // For atomic unrolled loops which are vector mapped, instigate more unrolling 2575 cl->set_notpassed_slp(); 2576 if (cl->is_main_loop()) { 2577 // if vector resources are limited, do not allow additional unrolling, also 2578 // do not unroll more on pure vector loops which were not reduced so that we can 2579 // program the post loop to single iteration execution. 2580 if (FLOATPRESSURE > 8) { 2581 C->set_major_progress(); 2582 cl->mark_do_unroll_only(); 2583 } 2584 } 2585 2586 if (do_reserve_copy()) { 2587 if (can_process_post_loop) { 2588 // Now create the difference of trip and limit and use it as our mask index. 2589 // Note: We limited the unroll of the vectorized loop so that 2590 // only vlen-1 size iterations can remain to be mask programmed. 2591 Node *incr = cl->incr(); 2592 SubINode *index = new SubINode(cl->limit(), cl->init_trip()); 2593 _igvn.register_new_node_with_optimizer(index); 2594 SetVectMaskINode *mask = new SetVectMaskINode(_phase->get_ctrl(cl->init_trip()), index); 2595 _igvn.register_new_node_with_optimizer(mask); 2596 // make this a single iteration loop 2597 AddINode *new_incr = new AddINode(incr->in(1), mask); 2598 _igvn.register_new_node_with_optimizer(new_incr); 2599 _phase->set_ctrl(new_incr, _phase->get_ctrl(incr)); 2600 _igvn.replace_node(incr, new_incr); 2601 cl->mark_is_multiversioned(); 2602 cl->loopexit()->add_flag(Node::Flag_has_vector_mask_set); 2603 } 2604 } 2605 } 2606 } 2607 } 2608 2609 if (do_reserve_copy()) { 2610 make_reversable.use_new(); 2611 } 2612 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);}) 2613 return; 2614 } 2615 2616 //------------------------------vector_opd--------------------------- 2617 // Create a vector operand for the nodes in pack p for operand: in(opd_idx) 2618 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) { 2619 Node* p0 = p->at(0); 2620 uint vlen = p->size(); 2621 Node* opd = p0->in(opd_idx); 2622 CountedLoopNode *cl = lpt()->_head->as_CountedLoop(); 2623 2624 if (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()) { 2625 // override vlen with the main loops vector length 2626 vlen = cl->slp_max_unroll(); 2627 } 2628 2629 if (same_inputs(p, opd_idx)) { 2630 if (opd->is_Vector() || opd->is_LoadVector()) { 2631 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector"); 2632 if (opd_idx == 2 && VectorNode::is_shift(p0)) { 2633 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");}) 2634 return NULL; 2635 } 2636 return opd; // input is matching vector 2637 } 2638 if ((opd_idx == 2) && VectorNode::is_shift(p0)) { 2639 Compile* C = _phase->C; 2640 Node* cnt = opd; 2641 // Vector instructions do not mask shift count, do it here. 2642 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2643 const TypeInt* t = opd->find_int_type(); 2644 if (t != NULL && t->is_con()) { 2645 juint shift = t->get_con(); 2646 if (shift > mask) { // Unsigned cmp 2647 cnt = ConNode::make(TypeInt::make(shift & mask)); 2648 } 2649 } else { 2650 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2651 cnt = ConNode::make(TypeInt::make(mask)); 2652 _igvn.register_new_node_with_optimizer(cnt); 2653 cnt = new AndINode(opd, cnt); 2654 _igvn.register_new_node_with_optimizer(cnt); 2655 _phase->set_ctrl(cnt, _phase->get_ctrl(opd)); 2656 } 2657 assert(opd->bottom_type()->isa_int(), "int type only"); 2658 if (!opd->bottom_type()->isa_int()) { 2659 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");}) 2660 return NULL; 2661 } 2662 // Move non constant shift count into vector register. 2663 cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0)); 2664 } 2665 if (cnt != opd) { 2666 _igvn.register_new_node_with_optimizer(cnt); 2667 _phase->set_ctrl(cnt, _phase->get_ctrl(opd)); 2668 } 2669 return cnt; 2670 } 2671 assert(!opd->is_StoreVector(), "such vector is not expected here"); 2672 if (opd->is_StoreVector()) { 2673 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");}) 2674 return NULL; 2675 } 2676 // Convert scalar input to vector with the same number of elements as 2677 // p0's vector. Use p0's type because size of operand's container in 2678 // vector should match p0's size regardless operand's size. 2679 const Type* p0_t = velt_type(p0); 2680 VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t); 2681 2682 _igvn.register_new_node_with_optimizer(vn); 2683 _phase->set_ctrl(vn, _phase->get_ctrl(opd)); 2684 #ifdef ASSERT 2685 if (TraceNewVectors) { 2686 tty->print("new Vector node: "); 2687 vn->dump(); 2688 } 2689 #endif 2690 return vn; 2691 } 2692 2693 // Insert pack operation 2694 BasicType bt = velt_basic_type(p0); 2695 PackNode* pk = PackNode::make(opd, vlen, bt); 2696 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); ) 2697 2698 for (uint i = 1; i < vlen; i++) { 2699 Node* pi = p->at(i); 2700 Node* in = pi->in(opd_idx); 2701 assert(my_pack(in) == NULL, "Should already have been unpacked"); 2702 if (my_pack(in) != NULL) { 2703 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");}) 2704 return NULL; 2705 } 2706 assert(opd_bt == in->bottom_type()->basic_type(), "all same type"); 2707 pk->add_opd(in); 2708 if (VectorNode::is_muladds2i(pi)) { 2709 Node* in2 = pi->in(opd_idx + 2); 2710 assert(my_pack(in2) == NULL, "Should already have been unpacked"); 2711 if (my_pack(in2) != NULL) { 2712 NOT_PRODUCT(if (is_trace_loop_reverse() || TraceLoopOpts) { tty->print_cr("Should already have been unpacked"); }) 2713 return NULL; 2714 } 2715 assert(opd_bt == in2->bottom_type()->basic_type(), "all same type"); 2716 pk->add_opd(in2); 2717 } 2718 } 2719 _igvn.register_new_node_with_optimizer(pk); 2720 _phase->set_ctrl(pk, _phase->get_ctrl(opd)); 2721 #ifdef ASSERT 2722 if (TraceNewVectors) { 2723 tty->print("new Vector node: "); 2724 pk->dump(); 2725 } 2726 #endif 2727 return pk; 2728 } 2729 2730 //------------------------------insert_extracts--------------------------- 2731 // If a use of pack p is not a vector use, then replace the 2732 // use with an extract operation. 2733 void SuperWord::insert_extracts(Node_List* p) { 2734 if (p->at(0)->is_Store()) return; 2735 assert(_n_idx_list.is_empty(), "empty (node,index) list"); 2736 2737 // Inspect each use of each pack member. For each use that is 2738 // not a vector use, replace the use with an extract operation. 2739 2740 for (uint i = 0; i < p->size(); i++) { 2741 Node* def = p->at(i); 2742 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2743 Node* use = def->fast_out(j); 2744 for (uint k = 0; k < use->req(); k++) { 2745 Node* n = use->in(k); 2746 if (def == n) { 2747 Node_List* u_pk = my_pack(use); 2748 if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) { 2749 _n_idx_list.push(use, k); 2750 } 2751 } 2752 } 2753 } 2754 } 2755 2756 while (_n_idx_list.is_nonempty()) { 2757 Node* use = _n_idx_list.node(); 2758 int idx = _n_idx_list.index(); 2759 _n_idx_list.pop(); 2760 Node* def = use->in(idx); 2761 2762 if (def->is_reduction()) continue; 2763 2764 // Insert extract operation 2765 _igvn.hash_delete(def); 2766 int def_pos = alignment(def) / data_size(def); 2767 2768 Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def)); 2769 _igvn.register_new_node_with_optimizer(ex); 2770 _phase->set_ctrl(ex, _phase->get_ctrl(def)); 2771 _igvn.replace_input_of(use, idx, ex); 2772 _igvn._worklist.push(def); 2773 2774 bb_insert_after(ex, bb_idx(def)); 2775 set_velt_type(ex, velt_type(def)); 2776 } 2777 } 2778 2779 //------------------------------is_vector_use--------------------------- 2780 // Is use->in(u_idx) a vector use? 2781 bool SuperWord::is_vector_use(Node* use, int u_idx) { 2782 Node_List* u_pk = my_pack(use); 2783 if (u_pk == NULL) return false; 2784 if (use->is_reduction()) return true; 2785 Node* def = use->in(u_idx); 2786 Node_List* d_pk = my_pack(def); 2787 if (d_pk == NULL) { 2788 // check for scalar promotion 2789 Node* n = u_pk->at(0)->in(u_idx); 2790 for (uint i = 1; i < u_pk->size(); i++) { 2791 if (u_pk->at(i)->in(u_idx) != n) return false; 2792 } 2793 return true; 2794 } 2795 if (VectorNode::is_muladds2i(use)) { 2796 // MulAddS2I takes shorts and produces ints - hence the special checks 2797 // on alignment and size. 2798 if (u_pk->size() * 2 != d_pk->size()) { 2799 return false; 2800 } 2801 for (uint i = 0; i < MIN2(d_pk->size(), u_pk->size()); i++) { 2802 Node* ui = u_pk->at(i); 2803 Node* di = d_pk->at(i); 2804 if (alignment(ui) != alignment(di) * 2) { 2805 return false; 2806 } 2807 } 2808 return true; 2809 } 2810 if (u_pk->size() != d_pk->size()) 2811 return false; 2812 for (uint i = 0; i < u_pk->size(); i++) { 2813 Node* ui = u_pk->at(i); 2814 Node* di = d_pk->at(i); 2815 if (ui->in(u_idx) != di || alignment(ui) != alignment(di)) 2816 return false; 2817 } 2818 return true; 2819 } 2820 2821 //------------------------------construct_bb--------------------------- 2822 // Construct reverse postorder list of block members 2823 bool SuperWord::construct_bb() { 2824 Node* entry = bb(); 2825 2826 assert(_stk.length() == 0, "stk is empty"); 2827 assert(_block.length() == 0, "block is empty"); 2828 assert(_data_entry.length() == 0, "data_entry is empty"); 2829 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty"); 2830 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty"); 2831 2832 // Find non-control nodes with no inputs from within block, 2833 // create a temporary map from node _idx to bb_idx for use 2834 // by the visited and post_visited sets, 2835 // and count number of nodes in block. 2836 int bb_ct = 0; 2837 for (uint i = 0; i < lpt()->_body.size(); i++) { 2838 Node *n = lpt()->_body.at(i); 2839 set_bb_idx(n, i); // Create a temporary map 2840 if (in_bb(n)) { 2841 if (n->is_LoadStore() || n->is_MergeMem() || 2842 (n->is_Proj() && !n->as_Proj()->is_CFG())) { 2843 // Bailout if the loop has LoadStore, MergeMem or data Proj 2844 // nodes. Superword optimization does not work with them. 2845 return false; 2846 } 2847 bb_ct++; 2848 if (!n->is_CFG()) { 2849 bool found = false; 2850 for (uint j = 0; j < n->req(); j++) { 2851 Node* def = n->in(j); 2852 if (def && in_bb(def)) { 2853 found = true; 2854 break; 2855 } 2856 } 2857 if (!found) { 2858 assert(n != entry, "can't be entry"); 2859 _data_entry.push(n); 2860 } 2861 } 2862 } 2863 } 2864 2865 // Find memory slices (head and tail) 2866 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) { 2867 Node *n = lp()->fast_out(i); 2868 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) { 2869 Node* n_tail = n->in(LoopNode::LoopBackControl); 2870 if (n_tail != n->in(LoopNode::EntryControl)) { 2871 if (!n_tail->is_Mem()) { 2872 assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name()); 2873 return false; // Bailout 2874 } 2875 _mem_slice_head.push(n); 2876 _mem_slice_tail.push(n_tail); 2877 } 2878 } 2879 } 2880 2881 // Create an RPO list of nodes in block 2882 2883 visited_clear(); 2884 post_visited_clear(); 2885 2886 // Push all non-control nodes with no inputs from within block, then control entry 2887 for (int j = 0; j < _data_entry.length(); j++) { 2888 Node* n = _data_entry.at(j); 2889 visited_set(n); 2890 _stk.push(n); 2891 } 2892 visited_set(entry); 2893 _stk.push(entry); 2894 2895 // Do a depth first walk over out edges 2896 int rpo_idx = bb_ct - 1; 2897 int size; 2898 int reduction_uses = 0; 2899 while ((size = _stk.length()) > 0) { 2900 Node* n = _stk.top(); // Leave node on stack 2901 if (!visited_test_set(n)) { 2902 // forward arc in graph 2903 } else if (!post_visited_test(n)) { 2904 // cross or back arc 2905 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2906 Node *use = n->fast_out(i); 2907 if (in_bb(use) && !visited_test(use) && 2908 // Don't go around backedge 2909 (!use->is_Phi() || n == entry)) { 2910 if (use->is_reduction()) { 2911 // First see if we can map the reduction on the given system we are on, then 2912 // make a data entry operation for each reduction we see. 2913 BasicType bt = use->bottom_type()->basic_type(); 2914 if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) { 2915 reduction_uses++; 2916 } 2917 } 2918 _stk.push(use); 2919 } 2920 } 2921 if (_stk.length() == size) { 2922 // There were no additional uses, post visit node now 2923 _stk.pop(); // Remove node from stack 2924 assert(rpo_idx >= 0, ""); 2925 _block.at_put_grow(rpo_idx, n); 2926 rpo_idx--; 2927 post_visited_set(n); 2928 assert(rpo_idx >= 0 || _stk.is_empty(), ""); 2929 } 2930 } else { 2931 _stk.pop(); // Remove post-visited node from stack 2932 } 2933 }//while 2934 2935 int ii_current = -1; 2936 unsigned int load_idx = (unsigned int)-1; 2937 _ii_order.clear(); 2938 // Create real map of block indices for nodes 2939 for (int j = 0; j < _block.length(); j++) { 2940 Node* n = _block.at(j); 2941 set_bb_idx(n, j); 2942 if (_do_vector_loop && n->is_Load()) { 2943 if (ii_current == -1) { 2944 ii_current = _clone_map.gen(n->_idx); 2945 _ii_order.push(ii_current); 2946 load_idx = _clone_map.idx(n->_idx); 2947 } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) { 2948 ii_current = _clone_map.gen(n->_idx); 2949 _ii_order.push(ii_current); 2950 } 2951 } 2952 }//for 2953 2954 // Ensure extra info is allocated. 2955 initialize_bb(); 2956 2957 #ifndef PRODUCT 2958 if (_vector_loop_debug && _ii_order.length() > 0) { 2959 tty->print("SuperWord::construct_bb: List of generations: "); 2960 for (int jj = 0; jj < _ii_order.length(); ++jj) { 2961 tty->print(" %d:%d", jj, _ii_order.at(jj)); 2962 } 2963 tty->print_cr(" "); 2964 } 2965 if (TraceSuperWord) { 2966 print_bb(); 2967 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE"); 2968 for (int m = 0; m < _data_entry.length(); m++) { 2969 tty->print("%3d ", m); 2970 _data_entry.at(m)->dump(); 2971 } 2972 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE"); 2973 for (int m = 0; m < _mem_slice_head.length(); m++) { 2974 tty->print("%3d ", m); _mem_slice_head.at(m)->dump(); 2975 tty->print(" "); _mem_slice_tail.at(m)->dump(); 2976 } 2977 } 2978 #endif 2979 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found"); 2980 return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0); 2981 } 2982 2983 //------------------------------initialize_bb--------------------------- 2984 // Initialize per node info 2985 void SuperWord::initialize_bb() { 2986 Node* last = _block.at(_block.length() - 1); 2987 grow_node_info(bb_idx(last)); 2988 } 2989 2990 //------------------------------bb_insert_after--------------------------- 2991 // Insert n into block after pos 2992 void SuperWord::bb_insert_after(Node* n, int pos) { 2993 int n_pos = pos + 1; 2994 // Make room 2995 for (int i = _block.length() - 1; i >= n_pos; i--) { 2996 _block.at_put_grow(i+1, _block.at(i)); 2997 } 2998 for (int j = _node_info.length() - 1; j >= n_pos; j--) { 2999 _node_info.at_put_grow(j+1, _node_info.at(j)); 3000 } 3001 // Set value 3002 _block.at_put_grow(n_pos, n); 3003 _node_info.at_put_grow(n_pos, SWNodeInfo::initial); 3004 // Adjust map from node->_idx to _block index 3005 for (int i = n_pos; i < _block.length(); i++) { 3006 set_bb_idx(_block.at(i), i); 3007 } 3008 } 3009 3010 //------------------------------compute_max_depth--------------------------- 3011 // Compute max depth for expressions from beginning of block 3012 // Use to prune search paths during test for independence. 3013 void SuperWord::compute_max_depth() { 3014 int ct = 0; 3015 bool again; 3016 do { 3017 again = false; 3018 for (int i = 0; i < _block.length(); i++) { 3019 Node* n = _block.at(i); 3020 if (!n->is_Phi()) { 3021 int d_orig = depth(n); 3022 int d_in = 0; 3023 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) { 3024 Node* pred = preds.current(); 3025 if (in_bb(pred)) { 3026 d_in = MAX2(d_in, depth(pred)); 3027 } 3028 } 3029 if (d_in + 1 != d_orig) { 3030 set_depth(n, d_in + 1); 3031 again = true; 3032 } 3033 } 3034 } 3035 ct++; 3036 } while (again); 3037 3038 if (TraceSuperWord && Verbose) { 3039 tty->print_cr("compute_max_depth iterated: %d times", ct); 3040 } 3041 } 3042 3043 //-------------------------compute_vector_element_type----------------------- 3044 // Compute necessary vector element type for expressions 3045 // This propagates backwards a narrower integer type when the 3046 // upper bits of the value are not needed. 3047 // Example: char a,b,c; a = b + c; 3048 // Normally the type of the add is integer, but for packed character 3049 // operations the type of the add needs to be char. 3050 void SuperWord::compute_vector_element_type() { 3051 if (TraceSuperWord && Verbose) { 3052 tty->print_cr("\ncompute_velt_type:"); 3053 } 3054 3055 // Initial type 3056 for (int i = 0; i < _block.length(); i++) { 3057 Node* n = _block.at(i); 3058 set_velt_type(n, container_type(n)); 3059 } 3060 3061 // Propagate integer narrowed type backwards through operations 3062 // that don't depend on higher order bits 3063 for (int i = _block.length() - 1; i >= 0; i--) { 3064 Node* n = _block.at(i); 3065 // Only integer types need be examined 3066 const Type* vtn = velt_type(n); 3067 if (vtn->basic_type() == T_INT) { 3068 uint start, end; 3069 VectorNode::vector_operands(n, &start, &end); 3070 3071 for (uint j = start; j < end; j++) { 3072 Node* in = n->in(j); 3073 // Don't propagate through a memory 3074 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT && 3075 data_size(n) < data_size(in)) { 3076 bool same_type = true; 3077 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) { 3078 Node *use = in->fast_out(k); 3079 if (!in_bb(use) || !same_velt_type(use, n)) { 3080 same_type = false; 3081 break; 3082 } 3083 } 3084 if (same_type) { 3085 // For right shifts of small integer types (bool, byte, char, short) 3086 // we need precise information about sign-ness. Only Load nodes have 3087 // this information because Store nodes are the same for signed and 3088 // unsigned values. And any arithmetic operation after a load may 3089 // expand a value to signed Int so such right shifts can't be used 3090 // because vector elements do not have upper bits of Int. 3091 const Type* vt = vtn; 3092 if (VectorNode::is_shift(in)) { 3093 Node* load = in->in(1); 3094 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) { 3095 vt = velt_type(load); 3096 } else if (in->Opcode() != Op_LShiftI) { 3097 // Widen type to Int to avoid creation of right shift vector 3098 // (align + data_size(s1) check in stmts_can_pack() will fail). 3099 // Note, left shifts work regardless type. 3100 vt = TypeInt::INT; 3101 } 3102 } 3103 set_velt_type(in, vt); 3104 } 3105 } 3106 } 3107 } 3108 } 3109 #ifndef PRODUCT 3110 if (TraceSuperWord && Verbose) { 3111 for (int i = 0; i < _block.length(); i++) { 3112 Node* n = _block.at(i); 3113 velt_type(n)->dump(); 3114 tty->print("\t"); 3115 n->dump(); 3116 } 3117 } 3118 #endif 3119 } 3120 3121 //------------------------------memory_alignment--------------------------- 3122 // Alignment within a vector memory reference 3123 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) { 3124 #ifndef PRODUCT 3125 if(TraceSuperWord && Verbose) { 3126 tty->print("SuperWord::memory_alignment within a vector memory reference for %d: ", s->_idx); s->dump(); 3127 } 3128 #endif 3129 NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);) 3130 SWPointer p(s, this, NULL, false); 3131 if (!p.valid()) { 3132 NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");) 3133 return bottom_align; 3134 } 3135 int vw = get_vw_bytes_special(s); 3136 if (vw < 2) { 3137 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");) 3138 return bottom_align; // No vectors for this type 3139 } 3140 int offset = p.offset_in_bytes(); 3141 offset += iv_adjust*p.memory_size(); 3142 int off_rem = offset % vw; 3143 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw; 3144 if (TraceSuperWord && Verbose) { 3145 tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod); 3146 } 3147 return off_mod; 3148 } 3149 3150 //---------------------------container_type--------------------------- 3151 // Smallest type containing range of values 3152 const Type* SuperWord::container_type(Node* n) { 3153 if (n->is_Mem()) { 3154 BasicType bt = n->as_Mem()->memory_type(); 3155 if (n->is_Store() && (bt == T_CHAR)) { 3156 // Use T_SHORT type instead of T_CHAR for stored values because any 3157 // preceding arithmetic operation extends values to signed Int. 3158 bt = T_SHORT; 3159 } 3160 if (n->Opcode() == Op_LoadUB) { 3161 // Adjust type for unsigned byte loads, it is important for right shifts. 3162 // T_BOOLEAN is used because there is no basic type representing type 3163 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only 3164 // size (one byte) and sign is important. 3165 bt = T_BOOLEAN; 3166 } 3167 return Type::get_const_basic_type(bt); 3168 } 3169 const Type* t = _igvn.type(n); 3170 if (t->basic_type() == T_INT) { 3171 // A narrow type of arithmetic operations will be determined by 3172 // propagating the type of memory operations. 3173 return TypeInt::INT; 3174 } 3175 return t; 3176 } 3177 3178 bool SuperWord::same_velt_type(Node* n1, Node* n2) { 3179 const Type* vt1 = velt_type(n1); 3180 const Type* vt2 = velt_type(n2); 3181 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) { 3182 // Compare vectors element sizes for integer types. 3183 return data_size(n1) == data_size(n2); 3184 } 3185 return vt1 == vt2; 3186 } 3187 3188 //------------------------------in_packset--------------------------- 3189 // Are s1 and s2 in a pack pair and ordered as s1,s2? 3190 bool SuperWord::in_packset(Node* s1, Node* s2) { 3191 for (int i = 0; i < _packset.length(); i++) { 3192 Node_List* p = _packset.at(i); 3193 assert(p->size() == 2, "must be"); 3194 if (p->at(0) == s1 && p->at(p->size()-1) == s2) { 3195 return true; 3196 } 3197 } 3198 return false; 3199 } 3200 3201 //------------------------------in_pack--------------------------- 3202 // Is s in pack p? 3203 Node_List* SuperWord::in_pack(Node* s, Node_List* p) { 3204 for (uint i = 0; i < p->size(); i++) { 3205 if (p->at(i) == s) { 3206 return p; 3207 } 3208 } 3209 return NULL; 3210 } 3211 3212 //------------------------------remove_pack_at--------------------------- 3213 // Remove the pack at position pos in the packset 3214 void SuperWord::remove_pack_at(int pos) { 3215 Node_List* p = _packset.at(pos); 3216 for (uint i = 0; i < p->size(); i++) { 3217 Node* s = p->at(i); 3218 set_my_pack(s, NULL); 3219 } 3220 _packset.remove_at(pos); 3221 } 3222 3223 void SuperWord::packset_sort(int n) { 3224 // simple bubble sort so that we capitalize with O(n) when its already sorted 3225 while (n != 0) { 3226 bool swapped = false; 3227 for (int i = 1; i < n; i++) { 3228 Node_List* q_low = _packset.at(i-1); 3229 Node_List* q_i = _packset.at(i); 3230 3231 // only swap when we find something to swap 3232 if (alignment(q_low->at(0)) > alignment(q_i->at(0))) { 3233 Node_List* t = q_i; 3234 *(_packset.adr_at(i)) = q_low; 3235 *(_packset.adr_at(i-1)) = q_i; 3236 swapped = true; 3237 } 3238 } 3239 if (swapped == false) break; 3240 n--; 3241 } 3242 } 3243 3244 //------------------------------executed_first--------------------------- 3245 // Return the node executed first in pack p. Uses the RPO block list 3246 // to determine order. 3247 Node* SuperWord::executed_first(Node_List* p) { 3248 Node* n = p->at(0); 3249 int n_rpo = bb_idx(n); 3250 for (uint i = 1; i < p->size(); i++) { 3251 Node* s = p->at(i); 3252 int s_rpo = bb_idx(s); 3253 if (s_rpo < n_rpo) { 3254 n = s; 3255 n_rpo = s_rpo; 3256 } 3257 } 3258 return n; 3259 } 3260 3261 //------------------------------executed_last--------------------------- 3262 // Return the node executed last in pack p. 3263 Node* SuperWord::executed_last(Node_List* p) { 3264 Node* n = p->at(0); 3265 int n_rpo = bb_idx(n); 3266 for (uint i = 1; i < p->size(); i++) { 3267 Node* s = p->at(i); 3268 int s_rpo = bb_idx(s); 3269 if (s_rpo > n_rpo) { 3270 n = s; 3271 n_rpo = s_rpo; 3272 } 3273 } 3274 return n; 3275 } 3276 3277 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) { 3278 LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest; 3279 for (uint i = 0; i < p->size(); i++) { 3280 Node* n = p->at(i); 3281 assert(n->is_Load(), "only meaningful for loads"); 3282 if (!n->depends_only_on_test()) { 3283 dep = LoadNode::Pinned; 3284 } 3285 } 3286 return dep; 3287 } 3288 3289 3290 //----------------------------align_initial_loop_index--------------------------- 3291 // Adjust pre-loop limit so that in main loop, a load/store reference 3292 // to align_to_ref will be a position zero in the vector. 3293 // (iv + k) mod vector_align == 0 3294 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) { 3295 CountedLoopNode *main_head = lp()->as_CountedLoop(); 3296 assert(main_head->is_main_loop(), ""); 3297 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head); 3298 assert(pre_end != NULL, "we must have a correct pre-loop"); 3299 Node *pre_opaq1 = pre_end->limit(); 3300 assert(pre_opaq1->Opcode() == Op_Opaque1, ""); 3301 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 3302 Node *lim0 = pre_opaq->in(1); 3303 3304 // Where we put new limit calculations 3305 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 3306 3307 // Ensure the original loop limit is available from the 3308 // pre-loop Opaque1 node. 3309 Node *orig_limit = pre_opaq->original_loop_limit(); 3310 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, ""); 3311 3312 SWPointer align_to_ref_p(align_to_ref, this, NULL, false); 3313 assert(align_to_ref_p.valid(), "sanity"); 3314 3315 // Given: 3316 // lim0 == original pre loop limit 3317 // V == v_align (power of 2) 3318 // invar == extra invariant piece of the address expression 3319 // e == offset [ +/- invar ] 3320 // 3321 // When reassociating expressions involving '%' the basic rules are: 3322 // (a - b) % k == 0 => a % k == b % k 3323 // and: 3324 // (a + b) % k == 0 => a % k == (k - b) % k 3325 // 3326 // For stride > 0 && scale > 0, 3327 // Derive the new pre-loop limit "lim" such that the two constraints: 3328 // (1) lim = lim0 + N (where N is some positive integer < V) 3329 // (2) (e + lim) % V == 0 3330 // are true. 3331 // 3332 // Substituting (1) into (2), 3333 // (e + lim0 + N) % V == 0 3334 // solve for N: 3335 // N = (V - (e + lim0)) % V 3336 // substitute back into (1), so that new limit 3337 // lim = lim0 + (V - (e + lim0)) % V 3338 // 3339 // For stride > 0 && scale < 0 3340 // Constraints: 3341 // lim = lim0 + N 3342 // (e - lim) % V == 0 3343 // Solving for lim: 3344 // (e - lim0 - N) % V == 0 3345 // N = (e - lim0) % V 3346 // lim = lim0 + (e - lim0) % V 3347 // 3348 // For stride < 0 && scale > 0 3349 // Constraints: 3350 // lim = lim0 - N 3351 // (e + lim) % V == 0 3352 // Solving for lim: 3353 // (e + lim0 - N) % V == 0 3354 // N = (e + lim0) % V 3355 // lim = lim0 - (e + lim0) % V 3356 // 3357 // For stride < 0 && scale < 0 3358 // Constraints: 3359 // lim = lim0 - N 3360 // (e - lim) % V == 0 3361 // Solving for lim: 3362 // (e - lim0 + N) % V == 0 3363 // N = (V - (e - lim0)) % V 3364 // lim = lim0 - (V - (e - lim0)) % V 3365 3366 int vw = vector_width_in_bytes(align_to_ref); 3367 int stride = iv_stride(); 3368 int scale = align_to_ref_p.scale_in_bytes(); 3369 int elt_size = align_to_ref_p.memory_size(); 3370 int v_align = vw / elt_size; 3371 assert(v_align > 1, "sanity"); 3372 int offset = align_to_ref_p.offset_in_bytes() / elt_size; 3373 Node *offsn = _igvn.intcon(offset); 3374 3375 Node *e = offsn; 3376 if (align_to_ref_p.invar() != NULL) { 3377 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt) 3378 Node* log2_elt = _igvn.intcon(exact_log2(elt_size)); 3379 Node* invar = align_to_ref_p.invar(); 3380 if (_igvn.type(invar)->isa_long()) { 3381 // Computations are done % (vector width/element size) so it's 3382 // safe to simply convert invar to an int and loose the upper 32 3383 // bit half. 3384 invar = new ConvL2INode(invar); 3385 _igvn.register_new_node_with_optimizer(invar); 3386 } 3387 Node* aref = new URShiftINode(invar, log2_elt); 3388 _igvn.register_new_node_with_optimizer(aref); 3389 _phase->set_ctrl(aref, pre_ctrl); 3390 if (align_to_ref_p.negate_invar()) { 3391 e = new SubINode(e, aref); 3392 } else { 3393 e = new AddINode(e, aref); 3394 } 3395 _igvn.register_new_node_with_optimizer(e); 3396 _phase->set_ctrl(e, pre_ctrl); 3397 } 3398 if (vw > ObjectAlignmentInBytes) { 3399 // incorporate base e +/- base && Mask >>> log2(elt) 3400 Node* xbase = new CastP2XNode(NULL, align_to_ref_p.base()); 3401 _igvn.register_new_node_with_optimizer(xbase); 3402 #ifdef _LP64 3403 xbase = new ConvL2INode(xbase); 3404 _igvn.register_new_node_with_optimizer(xbase); 3405 #endif 3406 Node* mask = _igvn.intcon(vw-1); 3407 Node* masked_xbase = new AndINode(xbase, mask); 3408 _igvn.register_new_node_with_optimizer(masked_xbase); 3409 Node* log2_elt = _igvn.intcon(exact_log2(elt_size)); 3410 Node* bref = new URShiftINode(masked_xbase, log2_elt); 3411 _igvn.register_new_node_with_optimizer(bref); 3412 _phase->set_ctrl(bref, pre_ctrl); 3413 e = new AddINode(e, bref); 3414 _igvn.register_new_node_with_optimizer(e); 3415 _phase->set_ctrl(e, pre_ctrl); 3416 } 3417 3418 // compute e +/- lim0 3419 if (scale < 0) { 3420 e = new SubINode(e, lim0); 3421 } else { 3422 e = new AddINode(e, lim0); 3423 } 3424 _igvn.register_new_node_with_optimizer(e); 3425 _phase->set_ctrl(e, pre_ctrl); 3426 3427 if (stride * scale > 0) { 3428 // compute V - (e +/- lim0) 3429 Node* va = _igvn.intcon(v_align); 3430 e = new SubINode(va, e); 3431 _igvn.register_new_node_with_optimizer(e); 3432 _phase->set_ctrl(e, pre_ctrl); 3433 } 3434 // compute N = (exp) % V 3435 Node* va_msk = _igvn.intcon(v_align - 1); 3436 Node* N = new AndINode(e, va_msk); 3437 _igvn.register_new_node_with_optimizer(N); 3438 _phase->set_ctrl(N, pre_ctrl); 3439 3440 // substitute back into (1), so that new limit 3441 // lim = lim0 + N 3442 Node* lim; 3443 if (stride < 0) { 3444 lim = new SubINode(lim0, N); 3445 } else { 3446 lim = new AddINode(lim0, N); 3447 } 3448 _igvn.register_new_node_with_optimizer(lim); 3449 _phase->set_ctrl(lim, pre_ctrl); 3450 Node* constrained = 3451 (stride > 0) ? (Node*) new MinINode(lim, orig_limit) 3452 : (Node*) new MaxINode(lim, orig_limit); 3453 _igvn.register_new_node_with_optimizer(constrained); 3454 _phase->set_ctrl(constrained, pre_ctrl); 3455 _igvn.replace_input_of(pre_opaq, 1, constrained); 3456 } 3457 3458 //----------------------------get_pre_loop_end--------------------------- 3459 // Find pre loop end from main loop. Returns null if none. 3460 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) { 3461 // The loop cannot be optimized if the graph shape at 3462 // the loop entry is inappropriate. 3463 if (!PhaseIdealLoop::is_canonical_loop_entry(cl)) { 3464 return NULL; 3465 } 3466 3467 Node* p_f = cl->skip_predicates()->in(0)->in(0); 3468 if (!p_f->is_IfFalse()) return NULL; 3469 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL; 3470 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd(); 3471 CountedLoopNode* loop_node = pre_end->loopnode(); 3472 if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL; 3473 return pre_end; 3474 } 3475 3476 //------------------------------init--------------------------- 3477 void SuperWord::init() { 3478 _dg.init(); 3479 _packset.clear(); 3480 _disjoint_ptrs.clear(); 3481 _block.clear(); 3482 _post_block.clear(); 3483 _data_entry.clear(); 3484 _mem_slice_head.clear(); 3485 _mem_slice_tail.clear(); 3486 _iteration_first.clear(); 3487 _iteration_last.clear(); 3488 _node_info.clear(); 3489 _align_to_ref = NULL; 3490 _lpt = NULL; 3491 _lp = NULL; 3492 _bb = NULL; 3493 _iv = NULL; 3494 _race_possible = 0; 3495 _early_return = false; 3496 _num_work_vecs = 0; 3497 _num_reductions = 0; 3498 } 3499 3500 //------------------------------restart--------------------------- 3501 void SuperWord::restart() { 3502 _dg.init(); 3503 _packset.clear(); 3504 _disjoint_ptrs.clear(); 3505 _block.clear(); 3506 _post_block.clear(); 3507 _data_entry.clear(); 3508 _mem_slice_head.clear(); 3509 _mem_slice_tail.clear(); 3510 _node_info.clear(); 3511 } 3512 3513 //------------------------------print_packset--------------------------- 3514 void SuperWord::print_packset() { 3515 #ifndef PRODUCT 3516 tty->print_cr("packset"); 3517 for (int i = 0; i < _packset.length(); i++) { 3518 tty->print_cr("Pack: %d", i); 3519 Node_List* p = _packset.at(i); 3520 print_pack(p); 3521 } 3522 #endif 3523 } 3524 3525 //------------------------------print_pack--------------------------- 3526 void SuperWord::print_pack(Node_List* p) { 3527 for (uint i = 0; i < p->size(); i++) { 3528 print_stmt(p->at(i)); 3529 } 3530 } 3531 3532 //------------------------------print_bb--------------------------- 3533 void SuperWord::print_bb() { 3534 #ifndef PRODUCT 3535 tty->print_cr("\nBlock"); 3536 for (int i = 0; i < _block.length(); i++) { 3537 Node* n = _block.at(i); 3538 tty->print("%d ", i); 3539 if (n) { 3540 n->dump(); 3541 } 3542 } 3543 #endif 3544 } 3545 3546 //------------------------------print_stmt--------------------------- 3547 void SuperWord::print_stmt(Node* s) { 3548 #ifndef PRODUCT 3549 tty->print(" align: %d \t", alignment(s)); 3550 s->dump(); 3551 #endif 3552 } 3553 3554 //------------------------------blank--------------------------- 3555 char* SuperWord::blank(uint depth) { 3556 static char blanks[101]; 3557 assert(depth < 101, "too deep"); 3558 for (uint i = 0; i < depth; i++) blanks[i] = ' '; 3559 blanks[depth] = '\0'; 3560 return blanks; 3561 } 3562 3563 3564 //==============================SWPointer=========================== 3565 #ifndef PRODUCT 3566 int SWPointer::Tracer::_depth = 0; 3567 #endif 3568 //----------------------------SWPointer------------------------ 3569 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) : 3570 _mem(mem), _slp(slp), _base(NULL), _adr(NULL), 3571 _scale(0), _offset(0), _invar(NULL), _negate_invar(false), 3572 _nstack(nstack), _analyze_only(analyze_only), 3573 _stack_idx(0) 3574 #ifndef PRODUCT 3575 , _tracer(slp) 3576 #endif 3577 { 3578 NOT_PRODUCT(_tracer.ctor_1(mem);) 3579 3580 Node* adr = mem->in(MemNode::Address); 3581 if (!adr->is_AddP()) { 3582 assert(!valid(), "too complex"); 3583 return; 3584 } 3585 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant) 3586 Node* base = adr->in(AddPNode::Base); 3587 // The base address should be loop invariant 3588 if (!invariant(base)) { 3589 assert(!valid(), "base address is loop variant"); 3590 return; 3591 } 3592 //unsafe reference could not be aligned appropriately without runtime checking 3593 if (base == NULL || base->bottom_type() == Type::TOP) { 3594 assert(!valid(), "unsafe access"); 3595 return; 3596 } 3597 3598 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();) 3599 NOT_PRODUCT(_tracer.ctor_2(adr);) 3600 3601 int i; 3602 for (i = 0; i < 3; i++) { 3603 NOT_PRODUCT(_tracer.ctor_3(adr, i);) 3604 3605 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) { 3606 assert(!valid(), "too complex"); 3607 return; 3608 } 3609 adr = adr->in(AddPNode::Address); 3610 NOT_PRODUCT(_tracer.ctor_4(adr, i);) 3611 3612 if (base == adr || !adr->is_AddP()) { 3613 NOT_PRODUCT(_tracer.ctor_5(adr, base, i);) 3614 break; // stop looking at addp's 3615 } 3616 } 3617 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();) 3618 NOT_PRODUCT(_tracer.ctor_6(mem);) 3619 3620 _base = base; 3621 _adr = adr; 3622 assert(valid(), "Usable"); 3623 } 3624 3625 // Following is used to create a temporary object during 3626 // the pattern match of an address expression. 3627 SWPointer::SWPointer(SWPointer* p) : 3628 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL), 3629 _scale(0), _offset(0), _invar(NULL), _negate_invar(false), 3630 _nstack(p->_nstack), _analyze_only(p->_analyze_only), 3631 _stack_idx(p->_stack_idx) 3632 #ifndef PRODUCT 3633 , _tracer(p->_slp) 3634 #endif 3635 {} 3636 3637 3638 bool SWPointer::invariant(Node* n) { 3639 NOT_PRODUCT(Tracer::Depth dd;) 3640 Node *n_c = phase()->get_ctrl(n); 3641 NOT_PRODUCT(_tracer.invariant_1(n, n_c);) 3642 return !lpt()->is_member(phase()->get_loop(n_c)); 3643 } 3644 //------------------------scaled_iv_plus_offset-------------------- 3645 // Match: k*iv + offset 3646 // where: k is a constant that maybe zero, and 3647 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional 3648 bool SWPointer::scaled_iv_plus_offset(Node* n) { 3649 NOT_PRODUCT(Tracer::Depth ddd;) 3650 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);) 3651 3652 if (scaled_iv(n)) { 3653 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);) 3654 return true; 3655 } 3656 3657 if (offset_plus_k(n)) { 3658 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);) 3659 return true; 3660 } 3661 3662 int opc = n->Opcode(); 3663 if (opc == Op_AddI) { 3664 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) { 3665 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);) 3666 return true; 3667 } 3668 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) { 3669 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);) 3670 return true; 3671 } 3672 } else if (opc == Op_SubI) { 3673 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) { 3674 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);) 3675 return true; 3676 } 3677 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) { 3678 _scale *= -1; 3679 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);) 3680 return true; 3681 } 3682 } 3683 3684 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);) 3685 return false; 3686 } 3687 3688 //----------------------------scaled_iv------------------------ 3689 // Match: k*iv where k is a constant that's not zero 3690 bool SWPointer::scaled_iv(Node* n) { 3691 NOT_PRODUCT(Tracer::Depth ddd;) 3692 NOT_PRODUCT(_tracer.scaled_iv_1(n);) 3693 3694 if (_scale != 0) { // already found a scale 3695 NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);) 3696 return false; 3697 } 3698 3699 if (n == iv()) { 3700 _scale = 1; 3701 NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);) 3702 return true; 3703 } 3704 if (_analyze_only && (invariant(n) == false)) { 3705 _nstack->push(n, _stack_idx++); 3706 } 3707 3708 int opc = n->Opcode(); 3709 if (opc == Op_MulI) { 3710 if (n->in(1) == iv() && n->in(2)->is_Con()) { 3711 _scale = n->in(2)->get_int(); 3712 NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);) 3713 return true; 3714 } else if (n->in(2) == iv() && n->in(1)->is_Con()) { 3715 _scale = n->in(1)->get_int(); 3716 NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);) 3717 return true; 3718 } 3719 } else if (opc == Op_LShiftI) { 3720 if (n->in(1) == iv() && n->in(2)->is_Con()) { 3721 _scale = 1 << n->in(2)->get_int(); 3722 NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);) 3723 return true; 3724 } 3725 } else if (opc == Op_ConvI2L) { 3726 if (n->in(1)->Opcode() == Op_CastII && 3727 n->in(1)->as_CastII()->has_range_check()) { 3728 // Skip range check dependent CastII nodes 3729 n = n->in(1); 3730 } 3731 if (scaled_iv_plus_offset(n->in(1))) { 3732 NOT_PRODUCT(_tracer.scaled_iv_7(n);) 3733 return true; 3734 } 3735 } else if (opc == Op_LShiftL) { 3736 if (!has_iv() && _invar == NULL) { 3737 // Need to preserve the current _offset value, so 3738 // create a temporary object for this expression subtree. 3739 // Hacky, so should re-engineer the address pattern match. 3740 NOT_PRODUCT(Tracer::Depth dddd;) 3741 SWPointer tmp(this); 3742 NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);) 3743 3744 if (tmp.scaled_iv_plus_offset(n->in(1))) { 3745 if (tmp._invar == NULL || _slp->do_vector_loop()) { 3746 int mult = 1 << n->in(2)->get_int(); 3747 _scale = tmp._scale * mult; 3748 _offset += tmp._offset * mult; 3749 NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);) 3750 return true; 3751 } 3752 } 3753 } 3754 } 3755 NOT_PRODUCT(_tracer.scaled_iv_10(n);) 3756 return false; 3757 } 3758 3759 //----------------------------offset_plus_k------------------------ 3760 // Match: offset is (k [+/- invariant]) 3761 // where k maybe zero and invariant is optional, but not both. 3762 bool SWPointer::offset_plus_k(Node* n, bool negate) { 3763 NOT_PRODUCT(Tracer::Depth ddd;) 3764 NOT_PRODUCT(_tracer.offset_plus_k_1(n);) 3765 3766 int opc = n->Opcode(); 3767 if (opc == Op_ConI) { 3768 _offset += negate ? -(n->get_int()) : n->get_int(); 3769 NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);) 3770 return true; 3771 } else if (opc == Op_ConL) { 3772 // Okay if value fits into an int 3773 const TypeLong* t = n->find_long_type(); 3774 if (t->higher_equal(TypeLong::INT)) { 3775 jlong loff = n->get_long(); 3776 jint off = (jint)loff; 3777 _offset += negate ? -off : loff; 3778 NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);) 3779 return true; 3780 } 3781 NOT_PRODUCT(_tracer.offset_plus_k_4(n);) 3782 return false; 3783 } 3784 if (_invar != NULL) { // already has an invariant 3785 NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);) 3786 return false; 3787 } 3788 3789 if (_analyze_only && (invariant(n) == false)) { 3790 _nstack->push(n, _stack_idx++); 3791 } 3792 if (opc == Op_AddI) { 3793 if (n->in(2)->is_Con() && invariant(n->in(1))) { 3794 _negate_invar = negate; 3795 _invar = n->in(1); 3796 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int(); 3797 NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);) 3798 return true; 3799 } else if (n->in(1)->is_Con() && invariant(n->in(2))) { 3800 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int(); 3801 _negate_invar = negate; 3802 _invar = n->in(2); 3803 NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);) 3804 return true; 3805 } 3806 } 3807 if (opc == Op_SubI) { 3808 if (n->in(2)->is_Con() && invariant(n->in(1))) { 3809 _negate_invar = negate; 3810 _invar = n->in(1); 3811 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int(); 3812 NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);) 3813 return true; 3814 } else if (n->in(1)->is_Con() && invariant(n->in(2))) { 3815 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int(); 3816 _negate_invar = !negate; 3817 _invar = n->in(2); 3818 NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);) 3819 return true; 3820 } 3821 } 3822 if (invariant(n)) { 3823 if (opc == Op_ConvI2L) { 3824 n = n->in(1); 3825 if (n->Opcode() == Op_CastII && 3826 n->as_CastII()->has_range_check()) { 3827 // Skip range check dependent CastII nodes 3828 assert(invariant(n), "sanity"); 3829 n = n->in(1); 3830 } 3831 } 3832 _negate_invar = negate; 3833 _invar = n; 3834 NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);) 3835 return true; 3836 } 3837 3838 NOT_PRODUCT(_tracer.offset_plus_k_11(n);) 3839 return false; 3840 } 3841 3842 //----------------------------print------------------------ 3843 void SWPointer::print() { 3844 #ifndef PRODUCT 3845 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n", 3846 _base != NULL ? _base->_idx : 0, 3847 _adr != NULL ? _adr->_idx : 0, 3848 _scale, _offset, 3849 _negate_invar?'-':'+', 3850 _invar != NULL ? _invar->_idx : 0); 3851 #endif 3852 } 3853 3854 //----------------------------tracing------------------------ 3855 #ifndef PRODUCT 3856 void SWPointer::Tracer::print_depth() { 3857 for (int ii = 0; ii<_depth; ++ii) tty->print(" "); 3858 } 3859 3860 void SWPointer::Tracer::ctor_1 (Node* mem) { 3861 if(_slp->is_trace_alignment()) { 3862 print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump(); 3863 } 3864 } 3865 3866 void SWPointer::Tracer::ctor_2(Node* adr) { 3867 if(_slp->is_trace_alignment()) { 3868 //store_depth(); 3869 inc_depth(); 3870 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump(); 3871 inc_depth(); 3872 print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump(); 3873 } 3874 } 3875 3876 void SWPointer::Tracer::ctor_3(Node* adr, int i) { 3877 if(_slp->is_trace_alignment()) { 3878 inc_depth(); 3879 Node* offset = adr->in(AddPNode::Offset); 3880 print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump(); 3881 } 3882 } 3883 3884 void SWPointer::Tracer::ctor_4(Node* adr, int i) { 3885 if(_slp->is_trace_alignment()) { 3886 inc_depth(); 3887 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump(); 3888 } 3889 } 3890 3891 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) { 3892 if(_slp->is_trace_alignment()) { 3893 inc_depth(); 3894 if (base == adr) { 3895 print_depth(); tty->print_cr(" \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i); 3896 } else if (!adr->is_AddP()) { 3897 print_depth(); tty->print_cr(" \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i); 3898 } 3899 } 3900 } 3901 3902 void SWPointer::Tracer::ctor_6(Node* mem) { 3903 if(_slp->is_trace_alignment()) { 3904 //restore_depth(); 3905 print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx); 3906 } 3907 } 3908 3909 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) { 3910 if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) { 3911 int is_member = _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)); 3912 int in_bb = _slp->in_bb(n); 3913 print_depth(); tty->print(" \\ "); tty->print_cr(" %d SWPointer::invariant conditions differ: n_c %d", n->_idx, n_c->_idx); 3914 print_depth(); tty->print(" \\ "); tty->print_cr("is_member %d, in_bb %d", is_member, in_bb); 3915 print_depth(); tty->print(" \\ "); n->dump(); 3916 print_depth(); tty->print(" \\ "); n_c->dump(); 3917 } 3918 } 3919 3920 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) { 3921 if(_slp->is_trace_alignment()) { 3922 print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx); 3923 n->dump(); 3924 } 3925 } 3926 3927 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) { 3928 if(_slp->is_trace_alignment()) { 3929 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx); 3930 } 3931 } 3932 3933 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) { 3934 if(_slp->is_trace_alignment()) { 3935 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx); 3936 } 3937 } 3938 3939 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) { 3940 if(_slp->is_trace_alignment()) { 3941 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx); 3942 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump(); 3943 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump(); 3944 } 3945 } 3946 3947 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) { 3948 if(_slp->is_trace_alignment()) { 3949 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx); 3950 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump(); 3951 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump(); 3952 } 3953 } 3954 3955 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) { 3956 if(_slp->is_trace_alignment()) { 3957 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx); 3958 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump(); 3959 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump(); 3960 } 3961 } 3962 3963 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) { 3964 if(_slp->is_trace_alignment()) { 3965 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx); 3966 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump(); 3967 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump(); 3968 } 3969 } 3970 3971 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) { 3972 if(_slp->is_trace_alignment()) { 3973 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx); 3974 } 3975 } 3976 3977 void SWPointer::Tracer::scaled_iv_1(Node* n) { 3978 if(_slp->is_trace_alignment()) { 3979 print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump(); 3980 } 3981 } 3982 3983 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) { 3984 if(_slp->is_trace_alignment()) { 3985 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx); 3986 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: _scale (%d) != 0", scale); 3987 } 3988 } 3989 3990 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) { 3991 if(_slp->is_trace_alignment()) { 3992 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale); 3993 } 3994 } 3995 3996 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) { 3997 if(_slp->is_trace_alignment()) { 3998 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale); 3999 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump(); 4000 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump(); 4001 } 4002 } 4003 4004 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) { 4005 if(_slp->is_trace_alignment()) { 4006 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale); 4007 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump(); 4008 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump(); 4009 } 4010 } 4011 4012 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) { 4013 if(_slp->is_trace_alignment()) { 4014 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale); 4015 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump(); 4016 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump(); 4017 } 4018 } 4019 4020 void SWPointer::Tracer::scaled_iv_7(Node* n) { 4021 if(_slp->is_trace_alignment()) { 4022 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx); 4023 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx); 4024 inc_depth(); inc_depth(); 4025 print_depth(); n->in(1)->dump(); 4026 dec_depth(); dec_depth(); 4027 } 4028 } 4029 4030 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) { 4031 if(_slp->is_trace_alignment()) { 4032 print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print(); 4033 } 4034 } 4035 4036 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) { 4037 if(_slp->is_trace_alignment()) { 4038 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset); 4039 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset, in(2) %d used to get mult = %d: _scale = %d, _offset = %d", 4040 n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset); 4041 inc_depth(); inc_depth(); 4042 print_depth(); n->in(1)->dump(); 4043 print_depth(); n->in(2)->dump(); 4044 dec_depth(); dec_depth(); 4045 } 4046 } 4047 4048 void SWPointer::Tracer::scaled_iv_10(Node* n) { 4049 if(_slp->is_trace_alignment()) { 4050 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx); 4051 } 4052 } 4053 4054 void SWPointer::Tracer::offset_plus_k_1(Node* n) { 4055 if(_slp->is_trace_alignment()) { 4056 print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump(); 4057 } 4058 } 4059 4060 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) { 4061 if(_slp->is_trace_alignment()) { 4062 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset); 4063 } 4064 } 4065 4066 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) { 4067 if(_slp->is_trace_alignment()) { 4068 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset); 4069 } 4070 } 4071 4072 void SWPointer::Tracer::offset_plus_k_4(Node* n) { 4073 if(_slp->is_trace_alignment()) { 4074 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx); 4075 print_depth(); tty->print_cr(" \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long()); 4076 } 4077 } 4078 4079 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) { 4080 if(_slp->is_trace_alignment()) { 4081 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx); 4082 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump(); 4083 } 4084 } 4085 4086 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) { 4087 if(_slp->is_trace_alignment()) { 4088 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d", 4089 n->_idx, _negate_invar, _invar->_idx, _offset); 4090 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump(); 4091 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump(); 4092 } 4093 } 4094 4095 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) { 4096 if(_slp->is_trace_alignment()) { 4097 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d", 4098 n->_idx, _negate_invar, _invar->_idx, _offset); 4099 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump(); 4100 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump(); 4101 } 4102 } 4103 4104 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) { 4105 if(_slp->is_trace_alignment()) { 4106 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d", 4107 n->_idx, _negate_invar, _invar->_idx, _offset); 4108 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump(); 4109 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump(); 4110 } 4111 } 4112 4113 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) { 4114 if(_slp->is_trace_alignment()) { 4115 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d", n->_idx, _negate_invar, _invar->_idx, _offset); 4116 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump(); 4117 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump(); 4118 } 4119 } 4120 4121 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) { 4122 if(_slp->is_trace_alignment()) { 4123 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d", n->_idx, _negate_invar, _invar->_idx, _offset); 4124 print_depth(); tty->print_cr(" \\ %d SWPointer::offset_plus_k: is invariant", n->_idx); 4125 } 4126 } 4127 4128 void SWPointer::Tracer::offset_plus_k_11(Node* n) { 4129 if(_slp->is_trace_alignment()) { 4130 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx); 4131 } 4132 } 4133 4134 #endif 4135 // ========================= OrderedPair ===================== 4136 4137 const OrderedPair OrderedPair::initial; 4138 4139 // ========================= SWNodeInfo ===================== 4140 4141 const SWNodeInfo SWNodeInfo::initial; 4142 4143 4144 // ============================ DepGraph =========================== 4145 4146 //------------------------------make_node--------------------------- 4147 // Make a new dependence graph node for an ideal node. 4148 DepMem* DepGraph::make_node(Node* node) { 4149 DepMem* m = new (_arena) DepMem(node); 4150 if (node != NULL) { 4151 assert(_map.at_grow(node->_idx) == NULL, "one init only"); 4152 _map.at_put_grow(node->_idx, m); 4153 } 4154 return m; 4155 } 4156 4157 //------------------------------make_edge--------------------------- 4158 // Make a new dependence graph edge from dpred -> dsucc 4159 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) { 4160 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head()); 4161 dpred->set_out_head(e); 4162 dsucc->set_in_head(e); 4163 return e; 4164 } 4165 4166 // ========================== DepMem ======================== 4167 4168 //------------------------------in_cnt--------------------------- 4169 int DepMem::in_cnt() { 4170 int ct = 0; 4171 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++; 4172 return ct; 4173 } 4174 4175 //------------------------------out_cnt--------------------------- 4176 int DepMem::out_cnt() { 4177 int ct = 0; 4178 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++; 4179 return ct; 4180 } 4181 4182 //------------------------------print----------------------------- 4183 void DepMem::print() { 4184 #ifndef PRODUCT 4185 tty->print(" DepNode %d (", _node->_idx); 4186 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) { 4187 Node* pred = p->pred()->node(); 4188 tty->print(" %d", pred != NULL ? pred->_idx : 0); 4189 } 4190 tty->print(") ["); 4191 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) { 4192 Node* succ = s->succ()->node(); 4193 tty->print(" %d", succ != NULL ? succ->_idx : 0); 4194 } 4195 tty->print_cr(" ]"); 4196 #endif 4197 } 4198 4199 // =========================== DepEdge ========================= 4200 4201 //------------------------------DepPreds--------------------------- 4202 void DepEdge::print() { 4203 #ifndef PRODUCT 4204 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx); 4205 #endif 4206 } 4207 4208 // =========================== DepPreds ========================= 4209 // Iterator over predecessor edges in the dependence graph. 4210 4211 //------------------------------DepPreds--------------------------- 4212 DepPreds::DepPreds(Node* n, DepGraph& dg) { 4213 _n = n; 4214 _done = false; 4215 if (_n->is_Store() || _n->is_Load()) { 4216 _next_idx = MemNode::Address; 4217 _end_idx = n->req(); 4218 _dep_next = dg.dep(_n)->in_head(); 4219 } else if (_n->is_Mem()) { 4220 _next_idx = 0; 4221 _end_idx = 0; 4222 _dep_next = dg.dep(_n)->in_head(); 4223 } else { 4224 _next_idx = 1; 4225 _end_idx = _n->req(); 4226 _dep_next = NULL; 4227 } 4228 next(); 4229 } 4230 4231 //------------------------------next--------------------------- 4232 void DepPreds::next() { 4233 if (_dep_next != NULL) { 4234 _current = _dep_next->pred()->node(); 4235 _dep_next = _dep_next->next_in(); 4236 } else if (_next_idx < _end_idx) { 4237 _current = _n->in(_next_idx++); 4238 } else { 4239 _done = true; 4240 } 4241 } 4242 4243 // =========================== DepSuccs ========================= 4244 // Iterator over successor edges in the dependence graph. 4245 4246 //------------------------------DepSuccs--------------------------- 4247 DepSuccs::DepSuccs(Node* n, DepGraph& dg) { 4248 _n = n; 4249 _done = false; 4250 if (_n->is_Load()) { 4251 _next_idx = 0; 4252 _end_idx = _n->outcnt(); 4253 _dep_next = dg.dep(_n)->out_head(); 4254 } else if (_n->is_Mem() || (_n->is_Phi() && _n->bottom_type() == Type::MEMORY)) { 4255 _next_idx = 0; 4256 _end_idx = 0; 4257 _dep_next = dg.dep(_n)->out_head(); 4258 } else { 4259 _next_idx = 0; 4260 _end_idx = _n->outcnt(); 4261 _dep_next = NULL; 4262 } 4263 next(); 4264 } 4265 4266 //-------------------------------next--------------------------- 4267 void DepSuccs::next() { 4268 if (_dep_next != NULL) { 4269 _current = _dep_next->succ()->node(); 4270 _dep_next = _dep_next->next_out(); 4271 } else if (_next_idx < _end_idx) { 4272 _current = _n->raw_out(_next_idx++); 4273 } else { 4274 _done = true; 4275 } 4276 } 4277 4278 // 4279 // --------------------------------- vectorization/simd ----------------------------------- 4280 // 4281 bool SuperWord::same_origin_idx(Node* a, Node* b) const { 4282 return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx); 4283 } 4284 bool SuperWord::same_generation(Node* a, Node* b) const { 4285 return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx); 4286 } 4287 4288 Node* SuperWord::find_phi_for_mem_dep(LoadNode* ld) { 4289 assert(in_bb(ld), "must be in block"); 4290 if (_clone_map.gen(ld->_idx) == _ii_first) { 4291 #ifndef PRODUCT 4292 if (_vector_loop_debug) { 4293 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d", 4294 _clone_map.gen(ld->_idx)); 4295 } 4296 #endif 4297 return NULL; //we think that any ld in the first gen being vectorizable 4298 } 4299 4300 Node* mem = ld->in(MemNode::Memory); 4301 if (mem->outcnt() <= 1) { 4302 // we don't want to remove the only edge from mem node to load 4303 #ifndef PRODUCT 4304 if (_vector_loop_debug) { 4305 tty->print_cr("SuperWord::find_phi_for_mem_dep input node %d to load %d has no other outputs and edge mem->load cannot be removed", 4306 mem->_idx, ld->_idx); 4307 ld->dump(); 4308 mem->dump(); 4309 } 4310 #endif 4311 return NULL; 4312 } 4313 if (!in_bb(mem) || same_generation(mem, ld)) { 4314 #ifndef PRODUCT 4315 if (_vector_loop_debug) { 4316 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d", 4317 _clone_map.gen(mem->_idx)); 4318 } 4319 #endif 4320 return NULL; // does not depend on loop volatile node or depends on the same generation 4321 } 4322 4323 //otherwise first node should depend on mem-phi 4324 Node* first = first_node(ld); 4325 assert(first->is_Load(), "must be Load"); 4326 Node* phi = first->as_Load()->in(MemNode::Memory); 4327 if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) { 4328 #ifndef PRODUCT 4329 if (_vector_loop_debug) { 4330 tty->print_cr("SuperWord::find_phi_for_mem_dep load is not vectorizable node, since it's `first` does not take input from mem phi"); 4331 ld->dump(); 4332 first->dump(); 4333 } 4334 #endif 4335 return NULL; 4336 } 4337 4338 Node* tail = 0; 4339 for (int m = 0; m < _mem_slice_head.length(); m++) { 4340 if (_mem_slice_head.at(m) == phi) { 4341 tail = _mem_slice_tail.at(m); 4342 } 4343 } 4344 if (tail == 0) { //test that found phi is in the list _mem_slice_head 4345 #ifndef PRODUCT 4346 if (_vector_loop_debug) { 4347 tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head", 4348 ld->_idx, phi->_idx); 4349 ld->dump(); 4350 phi->dump(); 4351 } 4352 #endif 4353 return NULL; 4354 } 4355 4356 // now all conditions are met 4357 return phi; 4358 } 4359 4360 Node* SuperWord::first_node(Node* nd) { 4361 for (int ii = 0; ii < _iteration_first.length(); ii++) { 4362 Node* nnn = _iteration_first.at(ii); 4363 if (same_origin_idx(nnn, nd)) { 4364 #ifndef PRODUCT 4365 if (_vector_loop_debug) { 4366 tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)", 4367 nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx)); 4368 } 4369 #endif 4370 return nnn; 4371 } 4372 } 4373 4374 #ifndef PRODUCT 4375 if (_vector_loop_debug) { 4376 tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)", 4377 nd->_idx, _clone_map.idx(nd->_idx)); 4378 } 4379 #endif 4380 return 0; 4381 } 4382 4383 Node* SuperWord::last_node(Node* nd) { 4384 for (int ii = 0; ii < _iteration_last.length(); ii++) { 4385 Node* nnn = _iteration_last.at(ii); 4386 if (same_origin_idx(nnn, nd)) { 4387 #ifndef PRODUCT 4388 if (_vector_loop_debug) { 4389 tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d", 4390 _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx)); 4391 } 4392 #endif 4393 return nnn; 4394 } 4395 } 4396 return 0; 4397 } 4398 4399 int SuperWord::mark_generations() { 4400 Node *ii_err = NULL, *tail_err = NULL; 4401 for (int i = 0; i < _mem_slice_head.length(); i++) { 4402 Node* phi = _mem_slice_head.at(i); 4403 assert(phi->is_Phi(), "must be phi"); 4404 4405 Node* tail = _mem_slice_tail.at(i); 4406 if (_ii_last == -1) { 4407 tail_err = tail; 4408 _ii_last = _clone_map.gen(tail->_idx); 4409 } 4410 else if (_ii_last != _clone_map.gen(tail->_idx)) { 4411 #ifndef PRODUCT 4412 if (TraceSuperWord && Verbose) { 4413 tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes "); 4414 tail->dump(); 4415 tail_err->dump(); 4416 } 4417 #endif 4418 return -1; 4419 } 4420 4421 // find first iteration in the loop 4422 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 4423 Node* ii = phi->fast_out(i); 4424 if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi 4425 if (_ii_first == -1) { 4426 ii_err = ii; 4427 _ii_first = _clone_map.gen(ii->_idx); 4428 } else if (_ii_first != _clone_map.gen(ii->_idx)) { 4429 #ifndef PRODUCT 4430 if (TraceSuperWord && Verbose) { 4431 tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first); 4432 ii->dump(); 4433 if (ii_err!= 0) { 4434 ii_err->dump(); 4435 } 4436 } 4437 #endif 4438 return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized 4439 } 4440 } 4441 }//for (DUIterator_Fast imax, 4442 }//for (int i... 4443 4444 if (_ii_first == -1 || _ii_last == -1) { 4445 if (TraceSuperWord && Verbose) { 4446 tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong"); 4447 } 4448 return -1; // something vent wrong 4449 } 4450 // collect nodes in the first and last generations 4451 assert(_iteration_first.length() == 0, "_iteration_first must be empty"); 4452 assert(_iteration_last.length() == 0, "_iteration_last must be empty"); 4453 for (int j = 0; j < _block.length(); j++) { 4454 Node* n = _block.at(j); 4455 node_idx_t gen = _clone_map.gen(n->_idx); 4456 if ((signed)gen == _ii_first) { 4457 _iteration_first.push(n); 4458 } else if ((signed)gen == _ii_last) { 4459 _iteration_last.push(n); 4460 } 4461 } 4462 4463 // building order of iterations 4464 if (_ii_order.length() == 0 && ii_err != 0) { 4465 assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb"); 4466 Node* nd = ii_err; 4467 while(_clone_map.gen(nd->_idx) != _ii_last) { 4468 _ii_order.push(_clone_map.gen(nd->_idx)); 4469 bool found = false; 4470 for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) { 4471 Node* use = nd->fast_out(i); 4472 if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) { 4473 found = true; 4474 nd = use; 4475 break; 4476 } 4477 }//for 4478 4479 if (found == false) { 4480 if (TraceSuperWord && Verbose) { 4481 tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx); 4482 } 4483 _ii_order.clear(); 4484 return -1; 4485 } 4486 } //while 4487 _ii_order.push(_clone_map.gen(nd->_idx)); 4488 } 4489 4490 #ifndef PRODUCT 4491 if (_vector_loop_debug) { 4492 tty->print_cr("SuperWord::mark_generations"); 4493 tty->print_cr("First generation (%d) nodes:", _ii_first); 4494 for (int ii = 0; ii < _iteration_first.length(); ii++) _iteration_first.at(ii)->dump(); 4495 tty->print_cr("Last generation (%d) nodes:", _ii_last); 4496 for (int ii = 0; ii < _iteration_last.length(); ii++) _iteration_last.at(ii)->dump(); 4497 tty->print_cr(" "); 4498 4499 tty->print("SuperWord::List of generations: "); 4500 for (int jj = 0; jj < _ii_order.length(); ++jj) { 4501 tty->print("%d:%d ", jj, _ii_order.at(jj)); 4502 } 4503 tty->print_cr(" "); 4504 } 4505 #endif 4506 4507 return _ii_first; 4508 } 4509 4510 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) { 4511 assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes"); 4512 assert(same_origin_idx(gold, fix), "should be clones of the same node"); 4513 Node* gin1 = gold->in(1); 4514 Node* gin2 = gold->in(2); 4515 Node* fin1 = fix->in(1); 4516 Node* fin2 = fix->in(2); 4517 bool swapped = false; 4518 4519 if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) { 4520 if (same_origin_idx(gin1, fin1) && 4521 same_origin_idx(gin2, fin2)) { 4522 return true; // nothing to fix 4523 } 4524 if (same_origin_idx(gin1, fin2) && 4525 same_origin_idx(gin2, fin1)) { 4526 fix->swap_edges(1, 2); 4527 swapped = true; 4528 } 4529 } 4530 // at least one input comes from outside of bb 4531 if (gin1->_idx == fin1->_idx) { 4532 return true; // nothing to fix 4533 } 4534 if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx)) { //swapping is expensive, check condition first 4535 fix->swap_edges(1, 2); 4536 swapped = true; 4537 } 4538 4539 if (swapped) { 4540 #ifndef PRODUCT 4541 if (_vector_loop_debug) { 4542 tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx); 4543 } 4544 #endif 4545 return true; 4546 } 4547 4548 if (TraceSuperWord && Verbose) { 4549 tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx); 4550 } 4551 4552 return false; 4553 } 4554 4555 bool SuperWord::pack_parallel() { 4556 #ifndef PRODUCT 4557 if (_vector_loop_debug) { 4558 tty->print_cr("SuperWord::pack_parallel: START"); 4559 } 4560 #endif 4561 4562 _packset.clear(); 4563 4564 for (int ii = 0; ii < _iteration_first.length(); ii++) { 4565 Node* nd = _iteration_first.at(ii); 4566 if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) { 4567 Node_List* pk = new Node_List(); 4568 pk->push(nd); 4569 for (int gen = 1; gen < _ii_order.length(); ++gen) { 4570 for (int kk = 0; kk < _block.length(); kk++) { 4571 Node* clone = _block.at(kk); 4572 if (same_origin_idx(clone, nd) && 4573 _clone_map.gen(clone->_idx) == _ii_order.at(gen)) { 4574 if (nd->is_Add() || nd->is_Mul()) { 4575 fix_commutative_inputs(nd, clone); 4576 } 4577 pk->push(clone); 4578 if (pk->size() == 4) { 4579 _packset.append(pk); 4580 #ifndef PRODUCT 4581 if (_vector_loop_debug) { 4582 tty->print_cr("SuperWord::pack_parallel: added pack "); 4583 pk->dump(); 4584 } 4585 #endif 4586 if (_clone_map.gen(clone->_idx) != _ii_last) { 4587 pk = new Node_List(); 4588 } 4589 } 4590 break; 4591 } 4592 } 4593 }//for 4594 }//if 4595 }//for 4596 4597 #ifndef PRODUCT 4598 if (_vector_loop_debug) { 4599 tty->print_cr("SuperWord::pack_parallel: END"); 4600 } 4601 #endif 4602 4603 return true; 4604 } 4605 4606 bool SuperWord::hoist_loads_in_graph() { 4607 GrowableArray<Node*> loads; 4608 4609 #ifndef PRODUCT 4610 if (_vector_loop_debug) { 4611 tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length()); 4612 } 4613 #endif 4614 4615 for (int i = 0; i < _mem_slice_head.length(); i++) { 4616 Node* n = _mem_slice_head.at(i); 4617 if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) { 4618 if (TraceSuperWord && Verbose) { 4619 tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx); 4620 } 4621 continue; 4622 } 4623 4624 #ifndef PRODUCT 4625 if (_vector_loop_debug) { 4626 tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d = _mem_slice_head.at(%d);", n->_idx, i); 4627 } 4628 #endif 4629 4630 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4631 Node* ld = n->fast_out(i); 4632 if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) { 4633 for (int i = 0; i < _block.length(); i++) { 4634 Node* ld2 = _block.at(i); 4635 if (ld2->is_Load() && same_origin_idx(ld, ld2) && 4636 !same_generation(ld, ld2)) { // <= do not collect the first generation ld 4637 #ifndef PRODUCT 4638 if (_vector_loop_debug) { 4639 tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)", 4640 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx); 4641 } 4642 #endif 4643 // could not do on-the-fly, since iterator is immutable 4644 loads.push(ld2); 4645 } 4646 }// for 4647 }//if 4648 }//for (DUIterator_Fast imax, 4649 }//for (int i = 0; i 4650 4651 for (int i = 0; i < loads.length(); i++) { 4652 LoadNode* ld = loads.at(i)->as_Load(); 4653 Node* phi = find_phi_for_mem_dep(ld); 4654 if (phi != NULL) { 4655 #ifndef PRODUCT 4656 if (_vector_loop_debug) { 4657 tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d", 4658 MemNode::Memory, ld->_idx, phi->_idx); 4659 } 4660 #endif 4661 _igvn.replace_input_of(ld, MemNode::Memory, phi); 4662 } 4663 }//for 4664 4665 restart(); // invalidate all basic structures, since we rebuilt the graph 4666 4667 if (TraceSuperWord && Verbose) { 4668 tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild"); 4669 } 4670 4671 return true; 4672 }