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