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