1 /*
2 * Copyright (c) 2007, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 */
23
24 #include "precompiled.hpp"
25 #include "compiler/compileLog.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/convertnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/matcher.hpp"
35 #include "opto/memnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/opcodes.hpp"
38 #include "opto/opaquenode.hpp"
39 #include "opto/superword.hpp"
40 #include "opto/vectornode.hpp"
41 #include "opto/movenode.hpp"
42 #include "utilities/powerOfTwo.hpp"
43
44 //
45 // S U P E R W O R D T R A N S F O R M
46 //=============================================================================
47
48 //------------------------------SuperWord---------------------------
49 SuperWord::SuperWord(PhaseIdealLoop* phase) :
50 _phase(phase),
51 _arena(phase->C->comp_arena()),
52 _igvn(phase->_igvn),
53 _packset(arena(), 8, 0, NULL), // packs for the current block
54 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
55 _block(arena(), 8, 0, NULL), // nodes in current block
56 _post_block(arena(), 8, 0, NULL), // nodes common to current block which are marked as post loop vectorizable
57 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
58 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
59 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
60 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
61 _clone_map(phase->C->clone_map()), // map of nodes created in cloning
62 _cmovev_kit(_arena, this), // map to facilitate CMoveV creation
63 _align_to_ref(NULL), // memory reference to align vectors to
64 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
65 _dg(_arena), // dependence graph
66 _visited(arena()), // visited node set
67 _post_visited(arena()), // post visited node set
68 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
69 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
70 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
71 _lpt(NULL), // loop tree node
72 _lp(NULL), // LoopNode
73 _bb(NULL), // basic block
74 _iv(NULL), // induction var
75 _race_possible(false), // cases where SDMU is true
76 _early_return(true), // analysis evaluations routine
77 _do_vector_loop(phase->C->do_vector_loop()), // whether to do vectorization/simd style
78 _do_reserve_copy(DoReserveCopyInSuperWord),
79 _num_work_vecs(0), // amount of vector work we have
80 _num_reductions(0), // amount of reduction work we have
81 _ii_first(-1), // first loop generation index - only if do_vector_loop()
82 _ii_last(-1), // last loop generation index - only if do_vector_loop()
83 _ii_order(arena(), 8, 0, 0)
84 {
85 #ifndef PRODUCT
86 _vector_loop_debug = 0;
87 if (_phase->C->method() != NULL) {
88 _vector_loop_debug = phase->C->directive()->VectorizeDebugOption;
89 }
90
91 #endif
92 }
93
94 //------------------------------transform_loop---------------------------
95 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
96 assert(UseSuperWord, "should be");
97 // Do vectors exist on this architecture?
98 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
99
100 assert(lpt->_head->is_CountedLoop(), "must be");
101 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
102
103 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
104
105 bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
106 if (post_loop_allowed) {
107 if (cl->is_reduction_loop()) return; // no predication mapping
108 Node *limit = cl->limit();
109 if (limit->is_Con()) return; // non constant limits only
110 // Now check the limit for expressions we do not handle
111 if (limit->is_Add()) {
112 Node *in2 = limit->in(2);
113 if (in2->is_Con()) {
114 int val = in2->get_int();
115 // should not try to program these cases
116 if (val < 0) return;
117 }
118 }
119 }
120
121 // skip any loop that has not been assigned max unroll by analysis
122 if (do_optimization) {
123 if (SuperWordLoopUnrollAnalysis && cl->slp_max_unroll() == 0) return;
124 }
125
126 // Check for no control flow in body (other than exit)
127 Node *cl_exit = cl->loopexit();
128 if (cl->is_main_loop() && (cl_exit->in(0) != lpt->_head)) {
129 #ifndef PRODUCT
130 if (TraceSuperWord) {
131 tty->print_cr("SuperWord::transform_loop: loop too complicated, cl_exit->in(0) != lpt->_head");
132 tty->print("cl_exit %d", cl_exit->_idx); cl_exit->dump();
133 tty->print("cl_exit->in(0) %d", cl_exit->in(0)->_idx); cl_exit->in(0)->dump();
134 tty->print("lpt->_head %d", lpt->_head->_idx); lpt->_head->dump();
135 lpt->dump_head();
136 }
137 #endif
138 return;
139 }
140
141 // Make sure the are no extra control users of the loop backedge
142 if (cl->back_control()->outcnt() != 1) {
143 return;
144 }
145
146 // Skip any loops already optimized by slp
147 if (cl->is_vectorized_loop()) return;
148
149 if (cl->is_unroll_only()) return;
150
151 if (cl->is_main_loop()) {
152 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
153 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
154 if (pre_end == NULL) return;
155 Node *pre_opaq1 = pre_end->limit();
156 if (pre_opaq1->Opcode() != Op_Opaque1) return;
157 }
158
159 init(); // initialize data structures
160
161 set_lpt(lpt);
162 set_lp(cl);
163
164 // For now, define one block which is the entire loop body
165 set_bb(cl);
166
167 if (do_optimization) {
168 assert(_packset.length() == 0, "packset must be empty");
169 SLP_extract();
170 if (PostLoopMultiversioning && Matcher::has_predicated_vectors()) {
171 if (cl->is_vectorized_loop() && cl->is_main_loop() && !cl->is_reduction_loop()) {
172 IdealLoopTree *lpt_next = lpt->_next;
173 CountedLoopNode *cl_next = lpt_next->_head->as_CountedLoop();
174 _phase->has_range_checks(lpt_next);
175 if (cl_next->is_post_loop() && !cl_next->range_checks_present()) {
176 if (!cl_next->is_vectorized_loop()) {
177 int slp_max_unroll_factor = cl->slp_max_unroll();
178 cl_next->set_slp_max_unroll(slp_max_unroll_factor);
179 }
180 }
181 }
182 }
183 }
184 }
185
186 //------------------------------early unrolling analysis------------------------------
187 void SuperWord::unrolling_analysis(int &local_loop_unroll_factor) {
188 bool is_slp = true;
189 ResourceMark rm;
190 size_t ignored_size = lpt()->_body.size();
191 int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size);
192 Node_Stack nstack((int)ignored_size);
193 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
194 Node *cl_exit = cl->loopexit_or_null();
195 int rpo_idx = _post_block.length();
196
197 assert(rpo_idx == 0, "post loop block is empty");
198
199 // First clear the entries
200 for (uint i = 0; i < lpt()->_body.size(); i++) {
201 ignored_loop_nodes[i] = -1;
202 }
203
204 int max_vector = Matcher::max_vector_size(T_BYTE);
205 bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
206
207 // Process the loop, some/all of the stack entries will not be in order, ergo
208 // need to preprocess the ignored initial state before we process the loop
209 for (uint i = 0; i < lpt()->_body.size(); i++) {
210 Node* n = lpt()->_body.at(i);
211 if (n == cl->incr() ||
212 n->is_reduction() ||
213 n->is_AddP() ||
214 n->is_Cmp() ||
215 n->is_IfTrue() ||
216 n->is_CountedLoop() ||
217 (n == cl_exit)) {
218 ignored_loop_nodes[i] = n->_idx;
219 continue;
220 }
221
222 if (n->is_If()) {
223 IfNode *iff = n->as_If();
224 if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
225 if (lpt()->is_loop_exit(iff)) {
226 ignored_loop_nodes[i] = n->_idx;
227 continue;
228 }
229 }
230 }
231
232 if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) {
233 Node* n_tail = n->in(LoopNode::LoopBackControl);
234 if (n_tail != n->in(LoopNode::EntryControl)) {
235 if (!n_tail->is_Mem()) {
236 is_slp = false;
237 break;
238 }
239 }
240 }
241
242 // This must happen after check of phi/if
243 if (n->is_Phi() || n->is_If()) {
244 ignored_loop_nodes[i] = n->_idx;
245 continue;
246 }
247
248 if (n->is_LoadStore() || n->is_MergeMem() ||
249 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
250 is_slp = false;
251 break;
252 }
253
254 // Ignore nodes with non-primitive type.
255 BasicType bt;
256 if (n->is_Mem()) {
257 bt = n->as_Mem()->memory_type();
258 } else {
259 bt = n->bottom_type()->basic_type();
260 }
261 if (is_java_primitive(bt) == false) {
262 ignored_loop_nodes[i] = n->_idx;
263 continue;
264 }
265
266 if (n->is_Mem()) {
267 MemNode* current = n->as_Mem();
268 Node* adr = n->in(MemNode::Address);
269 Node* n_ctrl = _phase->get_ctrl(adr);
270
271 // save a queue of post process nodes
272 if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) {
273 // Process the memory expression
274 int stack_idx = 0;
275 bool have_side_effects = true;
276 if (adr->is_AddP() == false) {
277 nstack.push(adr, stack_idx++);
278 } else {
279 // Mark the components of the memory operation in nstack
280 SWPointer p1(current, this, &nstack, true);
281 have_side_effects = p1.node_stack()->is_nonempty();
282 }
283
284 // Process the pointer stack
285 while (have_side_effects) {
286 Node* pointer_node = nstack.node();
287 for (uint j = 0; j < lpt()->_body.size(); j++) {
288 Node* cur_node = lpt()->_body.at(j);
289 if (cur_node == pointer_node) {
290 ignored_loop_nodes[j] = cur_node->_idx;
291 break;
292 }
293 }
294 nstack.pop();
295 have_side_effects = nstack.is_nonempty();
296 }
297 }
298 }
299 }
300
301 if (is_slp) {
302 // Now we try to find the maximum supported consistent vector which the machine
303 // description can use
304 bool small_basic_type = false;
305 bool flag_small_bt = false;
306 for (uint i = 0; i < lpt()->_body.size(); i++) {
307 if (ignored_loop_nodes[i] != -1) continue;
308
309 BasicType bt;
310 Node* n = lpt()->_body.at(i);
311 if (n->is_Mem()) {
312 bt = n->as_Mem()->memory_type();
313 } else {
314 bt = n->bottom_type()->basic_type();
315 }
316
317 if (post_loop_allowed) {
318 if (!small_basic_type) {
319 switch (bt) {
320 case T_CHAR:
321 case T_BYTE:
322 case T_SHORT:
323 small_basic_type = true;
324 break;
325
326 case T_LONG:
327 // TODO: Remove when support completed for mask context with LONG.
328 // Support needs to be augmented for logical qword operations, currently we map to dword
329 // buckets for vectors on logicals as these were legacy.
330 small_basic_type = true;
331 break;
332
333 default:
334 break;
335 }
336 }
337 }
338
339 if (is_java_primitive(bt) == false) continue;
340
341 int cur_max_vector = Matcher::max_vector_size(bt);
342
343 // If a max vector exists which is not larger than _local_loop_unroll_factor
344 // stop looking, we already have the max vector to map to.
345 if (cur_max_vector < local_loop_unroll_factor) {
346 is_slp = false;
347 if (TraceSuperWordLoopUnrollAnalysis) {
348 tty->print_cr("slp analysis fails: unroll limit greater than max vector\n");
349 }
350 break;
351 }
352
353 // Map the maximal common vector
354 if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
355 if (cur_max_vector < max_vector && !flag_small_bt) {
356 max_vector = cur_max_vector;
357 } else if (cur_max_vector > max_vector && UseSubwordForMaxVector) {
358 // Analyse subword in the loop to set maximum vector size to take advantage of full vector width for subword types.
359 // Here we analyze if narrowing is likely to happen and if it is we set vector size more aggressively.
360 // We check for possibility of narrowing by looking through chain operations using subword types.
361 if (is_subword_type(bt)) {
362 uint start, end;
363 VectorNode::vector_operands(n, &start, &end);
364
365 for (uint j = start; j < end; j++) {
366 Node* in = n->in(j);
367 // Don't propagate through a memory
368 if (!in->is_Mem() && in_bb(in) && in->bottom_type()->basic_type() == T_INT) {
369 bool same_type = true;
370 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
371 Node *use = in->fast_out(k);
372 if (!in_bb(use) && use->bottom_type()->basic_type() != bt) {
373 same_type = false;
374 break;
375 }
376 }
377 if (same_type) {
378 max_vector = cur_max_vector;
379 flag_small_bt = true;
380 cl->mark_subword_loop();
381 }
382 }
383 }
384 }
385 }
386 // We only process post loops on predicated targets where we want to
387 // mask map the loop to a single iteration
388 if (post_loop_allowed) {
389 _post_block.at_put_grow(rpo_idx++, n);
390 }
391 }
392 }
393 if (is_slp) {
394 local_loop_unroll_factor = max_vector;
395 cl->mark_passed_slp();
396 }
397 cl->mark_was_slp();
398 if (cl->is_main_loop()) {
399 cl->set_slp_max_unroll(local_loop_unroll_factor);
400 } else if (post_loop_allowed) {
401 if (!small_basic_type) {
402 // avoid replication context for small basic types in programmable masked loops
403 cl->set_slp_max_unroll(local_loop_unroll_factor);
404 }
405 }
406 }
407 }
408
409 //------------------------------SLP_extract---------------------------
410 // Extract the superword level parallelism
411 //
412 // 1) A reverse post-order of nodes in the block is constructed. By scanning
413 // this list from first to last, all definitions are visited before their uses.
414 //
415 // 2) A point-to-point dependence graph is constructed between memory references.
416 // This simplies the upcoming "independence" checker.
417 //
418 // 3) The maximum depth in the node graph from the beginning of the block
419 // to each node is computed. This is used to prune the graph search
420 // in the independence checker.
421 //
422 // 4) For integer types, the necessary bit width is propagated backwards
423 // from stores to allow packed operations on byte, char, and short
424 // integers. This reverses the promotion to type "int" that javac
425 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
426 //
427 // 5) One of the memory references is picked to be an aligned vector reference.
428 // The pre-loop trip count is adjusted to align this reference in the
429 // unrolled body.
430 //
431 // 6) The initial set of pack pairs is seeded with memory references.
432 //
433 // 7) The set of pack pairs is extended by following use->def and def->use links.
434 //
435 // 8) The pairs are combined into vector sized packs.
436 //
437 // 9) Reorder the memory slices to co-locate members of the memory packs.
438 //
439 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
440 // inserting scalar promotion, vector creation from multiple scalars, and
441 // extraction of scalar values from vectors.
442 //
443 void SuperWord::SLP_extract() {
444
445 #ifndef PRODUCT
446 if (_do_vector_loop && TraceSuperWord) {
447 tty->print("SuperWord::SLP_extract\n");
448 tty->print("input loop\n");
449 _lpt->dump_head();
450 _lpt->dump();
451 for (uint i = 0; i < _lpt->_body.size(); i++) {
452 _lpt->_body.at(i)->dump();
453 }
454 }
455 #endif
456 // Ready the block
457 if (!construct_bb()) {
458 return; // Exit if no interesting nodes or complex graph.
459 }
460
461 // build _dg, _disjoint_ptrs
462 dependence_graph();
463
464 // compute function depth(Node*)
465 compute_max_depth();
466
467 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
468 bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
469 if (cl->is_main_loop()) {
470 if (_do_vector_loop) {
471 if (mark_generations() != -1) {
472 hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
473
474 if (!construct_bb()) {
475 return; // Exit if no interesting nodes or complex graph.
476 }
477 dependence_graph();
478 compute_max_depth();
479 }
480
481 #ifndef PRODUCT
482 if (TraceSuperWord) {
483 tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
484 _lpt->dump_head();
485 for (int j = 0; j < _block.length(); j++) {
486 Node* n = _block.at(j);
487 int d = depth(n);
488 for (int i = 0; i < d; i++) tty->print("%s", " ");
489 tty->print("%d :", d);
490 n->dump();
491 }
492 }
493 #endif
494 }
495
496 compute_vector_element_type();
497
498 // Attempt vectorization
499
500 find_adjacent_refs();
501
502 extend_packlist();
503
504 if (_do_vector_loop) {
505 if (_packset.length() == 0) {
506 if (TraceSuperWord) {
507 tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
508 }
509 pack_parallel();
510 }
511 }
512
513 combine_packs();
514
515 construct_my_pack_map();
516 if (UseVectorCmov) {
517 merge_packs_to_cmovd();
518 }
519
520 filter_packs();
521
522 schedule();
523 } else if (post_loop_allowed) {
524 int saved_mapped_unroll_factor = cl->slp_max_unroll();
525 if (saved_mapped_unroll_factor) {
526 int vector_mapped_unroll_factor = saved_mapped_unroll_factor;
527
528 // now reset the slp_unroll_factor so that we can check the analysis mapped
529 // what the vector loop was mapped to
530 cl->set_slp_max_unroll(0);
531
532 // do the analysis on the post loop
533 unrolling_analysis(vector_mapped_unroll_factor);
534
535 // if our analyzed loop is a canonical fit, start processing it
536 if (vector_mapped_unroll_factor == saved_mapped_unroll_factor) {
537 // now add the vector nodes to packsets
538 for (int i = 0; i < _post_block.length(); i++) {
539 Node* n = _post_block.at(i);
540 Node_List* singleton = new Node_List();
541 singleton->push(n);
542 _packset.append(singleton);
543 set_my_pack(n, singleton);
544 }
545
546 // map base types for vector usage
547 compute_vector_element_type();
548 } else {
549 return;
550 }
551 } else {
552 // for some reason we could not map the slp analysis state of the vectorized loop
553 return;
554 }
555 }
556
557 output();
558 }
559
560 //------------------------------find_adjacent_refs---------------------------
561 // Find the adjacent memory references and create pack pairs for them.
562 // This is the initial set of packs that will then be extended by
563 // following use->def and def->use links. The align positions are
564 // assigned relative to the reference "align_to_ref"
565 void SuperWord::find_adjacent_refs() {
566 // Get list of memory operations
567 Node_List memops;
568 for (int i = 0; i < _block.length(); i++) {
569 Node* n = _block.at(i);
570 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
571 is_java_primitive(n->as_Mem()->memory_type())) {
572 int align = memory_alignment(n->as_Mem(), 0);
573 if (align != bottom_align) {
574 memops.push(n);
575 }
576 }
577 }
578
579 Node_List align_to_refs;
580 int max_idx;
581 int best_iv_adjustment = 0;
582 MemNode* best_align_to_mem_ref = NULL;
583
584 while (memops.size() != 0) {
585 // Find a memory reference to align to.
586 MemNode* mem_ref = find_align_to_ref(memops, max_idx);
587 if (mem_ref == NULL) break;
588 align_to_refs.push(mem_ref);
589 int iv_adjustment = get_iv_adjustment(mem_ref);
590
591 if (best_align_to_mem_ref == NULL) {
592 // Set memory reference which is the best from all memory operations
593 // to be used for alignment. The pre-loop trip count is modified to align
594 // this reference to a vector-aligned address.
595 best_align_to_mem_ref = mem_ref;
596 best_iv_adjustment = iv_adjustment;
597 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
598 }
599
600 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
601 // Set alignment relative to "align_to_ref" for all related memory operations.
602 for (int i = memops.size() - 1; i >= 0; i--) {
603 MemNode* s = memops.at(i)->as_Mem();
604 if (isomorphic(s, mem_ref) &&
605 (!_do_vector_loop || same_origin_idx(s, mem_ref))) {
606 SWPointer p2(s, this, NULL, false);
607 if (p2.comparable(align_to_ref_p)) {
608 int align = memory_alignment(s, iv_adjustment);
609 set_alignment(s, align);
610 }
611 }
612 }
613
614 // Create initial pack pairs of memory operations for which
615 // alignment is set and vectors will be aligned.
616 bool create_pack = true;
617 if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
618 if (!Matcher::misaligned_vectors_ok() || AlignVector) {
619 int vw = vector_width(mem_ref);
620 int vw_best = vector_width(best_align_to_mem_ref);
621 if (vw > vw_best) {
622 // Do not vectorize a memory access with more elements per vector
623 // if unaligned memory access is not allowed because number of
624 // iterations in pre-loop will be not enough to align it.
625 create_pack = false;
626 } else {
627 SWPointer p2(best_align_to_mem_ref, this, NULL, false);
628 if (align_to_ref_p.invar() != p2.invar()) {
629 // Do not vectorize memory accesses with different invariants
630 // if unaligned memory accesses are not allowed.
631 create_pack = false;
632 }
633 }
634 }
635 } else {
636 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
637 // Can't allow vectorization of unaligned memory accesses with the
638 // same type since it could be overlapped accesses to the same array.
639 create_pack = false;
640 } else {
641 // Allow independent (different type) unaligned memory operations
642 // if HW supports them.
643 if (!Matcher::misaligned_vectors_ok() || AlignVector) {
644 create_pack = false;
645 } else {
646 // Check if packs of the same memory type but
647 // with a different alignment were created before.
648 for (uint i = 0; i < align_to_refs.size(); i++) {
649 MemNode* mr = align_to_refs.at(i)->as_Mem();
650 if (mr == mem_ref) {
651 // Skip when we are looking at same memory operation.
652 continue;
653 }
654 if (same_velt_type(mr, mem_ref) &&
655 memory_alignment(mr, iv_adjustment) != 0)
656 create_pack = false;
657 }
658 }
659 }
660 }
661 if (create_pack) {
662 for (uint i = 0; i < memops.size(); i++) {
663 Node* s1 = memops.at(i);
664 int align = alignment(s1);
665 if (align == top_align) continue;
666 for (uint j = 0; j < memops.size(); j++) {
667 Node* s2 = memops.at(j);
668 if (alignment(s2) == top_align) continue;
669 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
670 if (stmts_can_pack(s1, s2, align)) {
671 Node_List* pair = new Node_List();
672 pair->push(s1);
673 pair->push(s2);
674 if (!_do_vector_loop || same_origin_idx(s1, s2)) {
675 _packset.append(pair);
676 }
677 }
678 }
679 }
680 }
681 } else { // Don't create unaligned pack
682 // First, remove remaining memory ops of the same type from the list.
683 for (int i = memops.size() - 1; i >= 0; i--) {
684 MemNode* s = memops.at(i)->as_Mem();
685 if (same_velt_type(s, mem_ref)) {
686 memops.remove(i);
687 }
688 }
689
690 // Second, remove already constructed packs of the same type.
691 for (int i = _packset.length() - 1; i >= 0; i--) {
692 Node_List* p = _packset.at(i);
693 MemNode* s = p->at(0)->as_Mem();
694 if (same_velt_type(s, mem_ref)) {
695 remove_pack_at(i);
696 }
697 }
698
699 // If needed find the best memory reference for loop alignment again.
700 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
701 // Put memory ops from remaining packs back on memops list for
702 // the best alignment search.
703 uint orig_msize = memops.size();
704 for (int i = 0; i < _packset.length(); i++) {
705 Node_List* p = _packset.at(i);
706 MemNode* s = p->at(0)->as_Mem();
707 assert(!same_velt_type(s, mem_ref), "sanity");
708 memops.push(s);
709 }
710 best_align_to_mem_ref = find_align_to_ref(memops, max_idx);
711 if (best_align_to_mem_ref == NULL) {
712 if (TraceSuperWord) {
713 tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
714 }
715 // best_align_to_mem_ref will be used for adjusting the pre-loop limit in
716 // SuperWord::align_initial_loop_index. Find one with the biggest vector size,
717 // smallest data size and smallest iv offset from memory ops from remaining packs.
718 if (_packset.length() > 0) {
719 if (orig_msize == 0) {
720 best_align_to_mem_ref = memops.at(max_idx)->as_Mem();
721 } else {
722 for (uint i = 0; i < orig_msize; i++) {
723 memops.remove(0);
724 }
725 best_align_to_mem_ref = find_align_to_ref(memops, max_idx);
726 assert(best_align_to_mem_ref == NULL, "sanity");
727 best_align_to_mem_ref = memops.at(max_idx)->as_Mem();
728 }
729 assert(best_align_to_mem_ref != NULL, "sanity");
730 }
731 break;
732 }
733 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
734 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
735 // Restore list.
736 while (memops.size() > orig_msize)
737 (void)memops.pop();
738 }
739 } // unaligned memory accesses
740
741 // Remove used mem nodes.
742 for (int i = memops.size() - 1; i >= 0; i--) {
743 MemNode* m = memops.at(i)->as_Mem();
744 if (alignment(m) != top_align) {
745 memops.remove(i);
746 }
747 }
748
749 } // while (memops.size() != 0
750 set_align_to_ref(best_align_to_mem_ref);
751
752 if (TraceSuperWord) {
753 tty->print_cr("\nAfter find_adjacent_refs");
754 print_packset();
755 }
756 }
757
758 #ifndef PRODUCT
759 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
760 if (is_trace_adjacent()) {
761 tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
762 best_align_to_mem_ref->_idx, best_iv_adjustment);
763 best_align_to_mem_ref->dump();
764 }
765 }
766 #endif
767
768 //------------------------------find_align_to_ref---------------------------
769 // Find a memory reference to align the loop induction variable to.
770 // Looks first at stores then at loads, looking for a memory reference
771 // with the largest number of references similar to it.
772 MemNode* SuperWord::find_align_to_ref(Node_List &memops, int &idx) {
773 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
774
775 // Count number of comparable memory ops
776 for (uint i = 0; i < memops.size(); i++) {
777 MemNode* s1 = memops.at(i)->as_Mem();
778 SWPointer p1(s1, this, NULL, false);
779 // Discard if pre loop can't align this reference
780 if (!ref_is_alignable(p1)) {
781 *cmp_ct.adr_at(i) = 0;
782 continue;
783 }
784 for (uint j = i+1; j < memops.size(); j++) {
785 MemNode* s2 = memops.at(j)->as_Mem();
786 if (isomorphic(s1, s2)) {
787 SWPointer p2(s2, this, NULL, false);
788 if (p1.comparable(p2)) {
789 (*cmp_ct.adr_at(i))++;
790 (*cmp_ct.adr_at(j))++;
791 }
792 }
793 }
794 }
795
796 // Find Store (or Load) with the greatest number of "comparable" references,
797 // biggest vector size, smallest data size and smallest iv offset.
798 int max_ct = 0;
799 int max_vw = 0;
800 int max_idx = -1;
801 int min_size = max_jint;
802 int min_iv_offset = max_jint;
803 for (uint j = 0; j < memops.size(); j++) {
804 MemNode* s = memops.at(j)->as_Mem();
805 if (s->is_Store()) {
806 int vw = vector_width_in_bytes(s);
807 assert(vw > 1, "sanity");
808 SWPointer p(s, this, NULL, false);
809 if ( cmp_ct.at(j) > max_ct ||
810 (cmp_ct.at(j) == max_ct &&
811 ( vw > max_vw ||
812 (vw == max_vw &&
813 ( data_size(s) < min_size ||
814 (data_size(s) == min_size &&
815 p.offset_in_bytes() < min_iv_offset)))))) {
816 max_ct = cmp_ct.at(j);
817 max_vw = vw;
818 max_idx = j;
819 min_size = data_size(s);
820 min_iv_offset = p.offset_in_bytes();
821 }
822 }
823 }
824 // If no stores, look at loads
825 if (max_ct == 0) {
826 for (uint j = 0; j < memops.size(); j++) {
827 MemNode* s = memops.at(j)->as_Mem();
828 if (s->is_Load()) {
829 int vw = vector_width_in_bytes(s);
830 assert(vw > 1, "sanity");
831 SWPointer p(s, this, NULL, false);
832 if ( cmp_ct.at(j) > max_ct ||
833 (cmp_ct.at(j) == max_ct &&
834 ( vw > max_vw ||
835 (vw == max_vw &&
836 ( data_size(s) < min_size ||
837 (data_size(s) == min_size &&
838 p.offset_in_bytes() < min_iv_offset)))))) {
839 max_ct = cmp_ct.at(j);
840 max_vw = vw;
841 max_idx = j;
842 min_size = data_size(s);
843 min_iv_offset = p.offset_in_bytes();
844 }
845 }
846 }
847 }
848
849 #ifdef ASSERT
850 if (TraceSuperWord && Verbose) {
851 tty->print_cr("\nVector memops after find_align_to_ref");
852 for (uint i = 0; i < memops.size(); i++) {
853 MemNode* s = memops.at(i)->as_Mem();
854 s->dump();
855 }
856 }
857 #endif
858
859 idx = max_idx;
860 if (max_ct > 0) {
861 #ifdef ASSERT
862 if (TraceSuperWord) {
863 tty->print("\nVector align to node: ");
864 memops.at(max_idx)->as_Mem()->dump();
865 }
866 #endif
867 return memops.at(max_idx)->as_Mem();
868 }
869 return NULL;
870 }
871
872 //------------------span_works_for_memory_size-----------------------------
873 static bool span_works_for_memory_size(MemNode* mem, int span, int mem_size, int offset) {
874 bool span_matches_memory = false;
875 if ((mem_size == type2aelembytes(T_BYTE) || mem_size == type2aelembytes(T_SHORT))
876 && ABS(span) == type2aelembytes(T_INT)) {
877 // There is a mismatch on span size compared to memory.
878 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
879 Node* use = mem->fast_out(j);
880 if (!VectorNode::is_type_transition_to_int(use)) {
881 return false;
882 }
883 }
884 // If all uses transition to integer, it means that we can successfully align even on mismatch.
885 return true;
886 }
887 else {
888 span_matches_memory = ABS(span) == mem_size;
889 }
890 return span_matches_memory && (ABS(offset) % mem_size) == 0;
891 }
892
893 //------------------------------ref_is_alignable---------------------------
894 // Can the preloop align the reference to position zero in the vector?
895 bool SuperWord::ref_is_alignable(SWPointer& p) {
896 if (!p.has_iv()) {
897 return true; // no induction variable
898 }
899 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
900 assert(pre_end != NULL, "we must have a correct pre-loop");
901 assert(pre_end->stride_is_con(), "pre loop stride is constant");
902 int preloop_stride = pre_end->stride_con();
903
904 int span = preloop_stride * p.scale_in_bytes();
905 int mem_size = p.memory_size();
906 int offset = p.offset_in_bytes();
907 // Stride one accesses are alignable if offset is aligned to memory operation size.
908 // Offset can be unaligned when UseUnalignedAccesses is used.
909 if (span_works_for_memory_size(p.mem(), span, mem_size, offset)) {
910 return true;
911 }
912 // If the initial offset from start of the object is computable,
913 // check if the pre-loop can align the final offset accordingly.
914 //
915 // In other words: Can we find an i such that the offset
916 // after i pre-loop iterations is aligned to vw?
917 // (init_offset + pre_loop) % vw == 0 (1)
918 // where
919 // pre_loop = i * span
920 // is the number of bytes added to the offset by i pre-loop iterations.
921 //
922 // For this to hold we need pre_loop to increase init_offset by
923 // pre_loop = vw - (init_offset % vw)
924 //
925 // This is only possible if pre_loop is divisible by span because each
926 // pre-loop iteration increases the initial offset by 'span' bytes:
927 // (vw - (init_offset % vw)) % span == 0
928 //
929 int vw = vector_width_in_bytes(p.mem());
930 assert(vw > 1, "sanity");
931 Node* init_nd = pre_end->init_trip();
932 if (init_nd->is_Con() && p.invar() == NULL) {
933 int init = init_nd->bottom_type()->is_int()->get_con();
934 int init_offset = init * p.scale_in_bytes() + offset;
935 if (init_offset < 0) { // negative offset from object start?
936 return false; // may happen in dead loop
937 }
938 if (vw % span == 0) {
939 // If vm is a multiple of span, we use formula (1).
940 if (span > 0) {
941 return (vw - (init_offset % vw)) % span == 0;
942 } else {
943 assert(span < 0, "nonzero stride * scale");
944 return (init_offset % vw) % -span == 0;
945 }
946 } else if (span % vw == 0) {
947 // If span is a multiple of vw, we can simplify formula (1) to:
948 // (init_offset + i * span) % vw == 0
949 // =>
950 // (init_offset % vw) + ((i * span) % vw) == 0
951 // =>
952 // init_offset % vw == 0
953 //
954 // Because we add a multiple of vw to the initial offset, the final
955 // offset is a multiple of vw if and only if init_offset is a multiple.
956 //
957 return (init_offset % vw) == 0;
958 }
959 }
960 return false;
961 }
962 //---------------------------get_vw_bytes_special------------------------
963 int SuperWord::get_vw_bytes_special(MemNode* s) {
964 // Get the vector width in bytes.
965 int vw = vector_width_in_bytes(s);
966
967 // Check for special case where there is an MulAddS2I usage where short vectors are going to need combined.
968 BasicType btype = velt_basic_type(s);
969 if (type2aelembytes(btype) == 2) {
970 bool should_combine_adjacent = true;
971 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
972 Node* user = s->fast_out(i);
973 if (!VectorNode::is_muladds2i(user)) {
974 should_combine_adjacent = false;
975 }
976 }
977 if (should_combine_adjacent) {
978 vw = MIN2(Matcher::max_vector_size(btype)*type2aelembytes(btype), vw * 2);
979 }
980 }
981
982 return vw;
983 }
984
985 //---------------------------get_iv_adjustment---------------------------
986 // Calculate loop's iv adjustment for this memory ops.
987 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
988 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
989 int offset = align_to_ref_p.offset_in_bytes();
990 int scale = align_to_ref_p.scale_in_bytes();
991 int elt_size = align_to_ref_p.memory_size();
992 int vw = get_vw_bytes_special(mem_ref);
993 assert(vw > 1, "sanity");
994 int iv_adjustment;
995 if (scale != 0) {
996 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
997 // At least one iteration is executed in pre-loop by default. As result
998 // several iterations are needed to align memory operations in main-loop even
999 // if offset is 0.
1000 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
1001 assert(((ABS(iv_adjustment_in_bytes) % elt_size) == 0),
1002 "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
1003 iv_adjustment = iv_adjustment_in_bytes/elt_size;
1004 } else {
1005 // This memory op is not dependent on iv (scale == 0)
1006 iv_adjustment = 0;
1007 }
1008
1009 #ifndef PRODUCT
1010 if (TraceSuperWord) {
1011 tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
1012 mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
1013 mem_ref->dump();
1014 }
1015 #endif
1016 return iv_adjustment;
1017 }
1018
1019 //---------------------------dependence_graph---------------------------
1020 // Construct dependency graph.
1021 // Add dependence edges to load/store nodes for memory dependence
1022 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
1023 void SuperWord::dependence_graph() {
1024 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
1025 // First, assign a dependence node to each memory node
1026 for (int i = 0; i < _block.length(); i++ ) {
1027 Node *n = _block.at(i);
1028 if (n->is_Mem() || (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1029 _dg.make_node(n);
1030 }
1031 }
1032
1033 // For each memory slice, create the dependences
1034 for (int i = 0; i < _mem_slice_head.length(); i++) {
1035 Node* n = _mem_slice_head.at(i);
1036 Node* n_tail = _mem_slice_tail.at(i);
1037
1038 // Get slice in predecessor order (last is first)
1039 if (cl->is_main_loop()) {
1040 mem_slice_preds(n_tail, n, _nlist);
1041 }
1042
1043 #ifndef PRODUCT
1044 if(TraceSuperWord && Verbose) {
1045 tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
1046 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
1047 _nlist.at(j)->dump();
1048 }
1049 }
1050 #endif
1051 // Make the slice dependent on the root
1052 DepMem* slice = _dg.dep(n);
1053 _dg.make_edge(_dg.root(), slice);
1054
1055 // Create a sink for the slice
1056 DepMem* slice_sink = _dg.make_node(NULL);
1057 _dg.make_edge(slice_sink, _dg.tail());
1058
1059 // Now visit each pair of memory ops, creating the edges
1060 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
1061 Node* s1 = _nlist.at(j);
1062
1063 // If no dependency yet, use slice
1064 if (_dg.dep(s1)->in_cnt() == 0) {
1065 _dg.make_edge(slice, s1);
1066 }
1067 SWPointer p1(s1->as_Mem(), this, NULL, false);
1068 bool sink_dependent = true;
1069 for (int k = j - 1; k >= 0; k--) {
1070 Node* s2 = _nlist.at(k);
1071 if (s1->is_Load() && s2->is_Load())
1072 continue;
1073 SWPointer p2(s2->as_Mem(), this, NULL, false);
1074
1075 int cmp = p1.cmp(p2);
1076 if (SuperWordRTDepCheck &&
1077 p1.base() != p2.base() && p1.valid() && p2.valid()) {
1078 // Create a runtime check to disambiguate
1079 OrderedPair pp(p1.base(), p2.base());
1080 _disjoint_ptrs.append_if_missing(pp);
1081 } else if (!SWPointer::not_equal(cmp)) {
1082 // Possibly same address
1083 _dg.make_edge(s1, s2);
1084 sink_dependent = false;
1085 }
1086 }
1087 if (sink_dependent) {
1088 _dg.make_edge(s1, slice_sink);
1089 }
1090 }
1091
1092 if (TraceSuperWord) {
1093 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
1094 for (int q = 0; q < _nlist.length(); q++) {
1095 _dg.print(_nlist.at(q));
1096 }
1097 tty->cr();
1098 }
1099
1100 _nlist.clear();
1101 }
1102
1103 if (TraceSuperWord) {
1104 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
1105 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
1106 _disjoint_ptrs.at(r).print();
1107 tty->cr();
1108 }
1109 tty->cr();
1110 }
1111
1112 }
1113
1114 //---------------------------mem_slice_preds---------------------------
1115 // Return a memory slice (node list) in predecessor order starting at "start"
1116 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
1117 assert(preds.length() == 0, "start empty");
1118 Node* n = start;
1119 Node* prev = NULL;
1120 while (true) {
1121 NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
1122 assert(in_bb(n), "must be in block");
1123 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1124 Node* out = n->fast_out(i);
1125 if (out->is_Load()) {
1126 if (in_bb(out)) {
1127 preds.push(out);
1128 if (TraceSuperWord && Verbose) {
1129 tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
1130 }
1131 }
1132 } else {
1133 // FIXME
1134 if (out->is_MergeMem() && !in_bb(out)) {
1135 // Either unrolling is causing a memory edge not to disappear,
1136 // or need to run igvn.optimize() again before SLP
1137 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
1138 // Ditto. Not sure what else to check further.
1139 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
1140 // StoreCM has an input edge used as a precedence edge.
1141 // Maybe an issue when oop stores are vectorized.
1142 } else {
1143 assert(out == prev || prev == NULL, "no branches off of store slice");
1144 }
1145 }//else
1146 }//for
1147 if (n == stop) break;
1148 preds.push(n);
1149 if (TraceSuperWord && Verbose) {
1150 tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
1151 }
1152 prev = n;
1153 assert(n->is_Mem(), "unexpected node %s", n->Name());
1154 n = n->in(MemNode::Memory);
1155 }
1156 }
1157
1158 //------------------------------stmts_can_pack---------------------------
1159 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
1160 // s1 aligned at "align"
1161 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
1162
1163 // Do not use superword for non-primitives
1164 BasicType bt1 = velt_basic_type(s1);
1165 BasicType bt2 = velt_basic_type(s2);
1166 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
1167 return false;
1168 if (Matcher::max_vector_size(bt1) < 2) {
1169 return false; // No vectors for this type
1170 }
1171
1172 if (isomorphic(s1, s2)) {
1173 if ((independent(s1, s2) && have_similar_inputs(s1, s2)) || reduction(s1, s2)) {
1174 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
1175 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
1176 int s1_align = alignment(s1);
1177 int s2_align = alignment(s2);
1178 if (s1_align == top_align || s1_align == align) {
1179 if (s2_align == top_align || s2_align == align + data_size(s1)) {
1180 return true;
1181 }
1182 }
1183 }
1184 }
1185 }
1186 }
1187 return false;
1188 }
1189
1190 //------------------------------exists_at---------------------------
1191 // Does s exist in a pack at position pos?
1192 bool SuperWord::exists_at(Node* s, uint pos) {
1193 for (int i = 0; i < _packset.length(); i++) {
1194 Node_List* p = _packset.at(i);
1195 if (p->at(pos) == s) {
1196 return true;
1197 }
1198 }
1199 return false;
1200 }
1201
1202 //------------------------------are_adjacent_refs---------------------------
1203 // Is s1 immediately before s2 in memory?
1204 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
1205 if (!s1->is_Mem() || !s2->is_Mem()) return false;
1206 if (!in_bb(s1) || !in_bb(s2)) return false;
1207
1208 // Do not use superword for non-primitives
1209 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
1210 !is_java_primitive(s2->as_Mem()->memory_type())) {
1211 return false;
1212 }
1213
1214 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
1215 // only pack memops that are in the same alias set until that's fixed.
1216 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1217 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1218 return false;
1219 SWPointer p1(s1->as_Mem(), this, NULL, false);
1220 SWPointer p2(s2->as_Mem(), this, NULL, false);
1221 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1222 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1223 return diff == data_size(s1);
1224 }
1225
1226 //------------------------------isomorphic---------------------------
1227 // Are s1 and s2 similar?
1228 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1229 if (s1->Opcode() != s2->Opcode()) return false;
1230 if (s1->req() != s2->req()) return false;
1231 if (!same_velt_type(s1, s2)) return false;
1232 Node* s1_ctrl = s1->in(0);
1233 Node* s2_ctrl = s2->in(0);
1234 // If the control nodes are equivalent, no further checks are required to test for isomorphism.
1235 if (s1_ctrl == s2_ctrl) {
1236 return true;
1237 } else {
1238 bool s1_ctrl_inv = ((s1_ctrl == NULL) ? true : lpt()->is_invariant(s1_ctrl));
1239 bool s2_ctrl_inv = ((s2_ctrl == NULL) ? true : lpt()->is_invariant(s2_ctrl));
1240 // If the control nodes are not invariant for the loop, fail isomorphism test.
1241 if (!s1_ctrl_inv || !s2_ctrl_inv) {
1242 return false;
1243 }
1244 if(s1_ctrl != NULL && s2_ctrl != NULL) {
1245 if (s1_ctrl->is_Proj()) {
1246 s1_ctrl = s1_ctrl->in(0);
1247 assert(lpt()->is_invariant(s1_ctrl), "must be invariant");
1248 }
1249 if (s2_ctrl->is_Proj()) {
1250 s2_ctrl = s2_ctrl->in(0);
1251 assert(lpt()->is_invariant(s2_ctrl), "must be invariant");
1252 }
1253 if (!s1_ctrl->is_RangeCheck() || !s2_ctrl->is_RangeCheck()) {
1254 return false;
1255 }
1256 }
1257 // Control nodes are invariant. However, we have no way of checking whether they resolve
1258 // in an equivalent manner. But, we know that invariant range checks are guaranteed to
1259 // throw before the loop (if they would have thrown). Thus, the loop would not have been reached.
1260 // Therefore, if the control nodes for both are range checks, we accept them to be isomorphic.
1261 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1262 Node* t1 = s1->fast_out(i);
1263 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1264 Node* t2 = s2->fast_out(j);
1265 if (VectorNode::is_muladds2i(t1) && VectorNode::is_muladds2i(t2)) {
1266 return true;
1267 }
1268 }
1269 }
1270 }
1271 return false;
1272 }
1273
1274 //------------------------------independent---------------------------
1275 // Is there no data path from s1 to s2 or s2 to s1?
1276 bool SuperWord::independent(Node* s1, Node* s2) {
1277 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1278 int d1 = depth(s1);
1279 int d2 = depth(s2);
1280 if (d1 == d2) return s1 != s2;
1281 Node* deep = d1 > d2 ? s1 : s2;
1282 Node* shallow = d1 > d2 ? s2 : s1;
1283
1284 visited_clear();
1285
1286 return independent_path(shallow, deep);
1287 }
1288
1289 //--------------------------have_similar_inputs-----------------------
1290 // For a node pair (s1, s2) which is isomorphic and independent,
1291 // do s1 and s2 have similar input edges?
1292 bool SuperWord::have_similar_inputs(Node* s1, Node* s2) {
1293 // assert(isomorphic(s1, s2) == true, "check isomorphic");
1294 // assert(independent(s1, s2) == true, "check independent");
1295 if (s1->req() > 1 && !s1->is_Store() && !s1->is_Load()) {
1296 for (uint i = 1; i < s1->req(); i++) {
1297 if (s1->in(i)->Opcode() != s2->in(i)->Opcode()) return false;
1298 }
1299 }
1300 return true;
1301 }
1302
1303 //------------------------------reduction---------------------------
1304 // Is there a data path between s1 and s2 and the nodes reductions?
1305 bool SuperWord::reduction(Node* s1, Node* s2) {
1306 bool retValue = false;
1307 int d1 = depth(s1);
1308 int d2 = depth(s2);
1309 if (d1 + 1 == d2) {
1310 if (s1->is_reduction() && s2->is_reduction()) {
1311 // This is an ordered set, so s1 should define s2
1312 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1313 Node* t1 = s1->fast_out(i);
1314 if (t1 == s2) {
1315 // both nodes are reductions and connected
1316 retValue = true;
1317 }
1318 }
1319 }
1320 }
1321
1322 return retValue;
1323 }
1324
1325 //------------------------------independent_path------------------------------
1326 // Helper for independent
1327 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1328 if (dp >= 1000) return false; // stop deep recursion
1329 visited_set(deep);
1330 int shal_depth = depth(shallow);
1331 assert(shal_depth <= depth(deep), "must be");
1332 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1333 Node* pred = preds.current();
1334 if (in_bb(pred) && !visited_test(pred)) {
1335 if (shallow == pred) {
1336 return false;
1337 }
1338 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1339 return false;
1340 }
1341 }
1342 }
1343 return true;
1344 }
1345
1346 //------------------------------set_alignment---------------------------
1347 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1348 set_alignment(s1, align);
1349 if (align == top_align || align == bottom_align) {
1350 set_alignment(s2, align);
1351 } else {
1352 set_alignment(s2, align + data_size(s1));
1353 }
1354 }
1355
1356 //------------------------------data_size---------------------------
1357 int SuperWord::data_size(Node* s) {
1358 Node* use = NULL; //test if the node is a candidate for CMoveV optimization, then return the size of CMov
1359 if (UseVectorCmov) {
1360 use = _cmovev_kit.is_Bool_candidate(s);
1361 if (use != NULL) {
1362 return data_size(use);
1363 }
1364 use = _cmovev_kit.is_CmpD_candidate(s);
1365 if (use != NULL) {
1366 return data_size(use);
1367 }
1368 }
1369
1370 int bsize = type2aelembytes(velt_basic_type(s));
1371 assert(bsize != 0, "valid size");
1372 return bsize;
1373 }
1374
1375 //------------------------------extend_packlist---------------------------
1376 // Extend packset by following use->def and def->use links from pack members.
1377 void SuperWord::extend_packlist() {
1378 bool changed;
1379 do {
1380 packset_sort(_packset.length());
1381 changed = false;
1382 for (int i = 0; i < _packset.length(); i++) {
1383 Node_List* p = _packset.at(i);
1384 changed |= follow_use_defs(p);
1385 changed |= follow_def_uses(p);
1386 }
1387 } while (changed);
1388
1389 if (_race_possible) {
1390 for (int i = 0; i < _packset.length(); i++) {
1391 Node_List* p = _packset.at(i);
1392 order_def_uses(p);
1393 }
1394 }
1395
1396 if (TraceSuperWord) {
1397 tty->print_cr("\nAfter extend_packlist");
1398 print_packset();
1399 }
1400 }
1401
1402 //------------------------------follow_use_defs---------------------------
1403 // Extend the packset by visiting operand definitions of nodes in pack p
1404 bool SuperWord::follow_use_defs(Node_List* p) {
1405 assert(p->size() == 2, "just checking");
1406 Node* s1 = p->at(0);
1407 Node* s2 = p->at(1);
1408 assert(s1->req() == s2->req(), "just checking");
1409 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1410
1411 if (s1->is_Load()) return false;
1412
1413 int align = alignment(s1);
1414 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1415 bool changed = false;
1416 int start = s1->is_Store() ? MemNode::ValueIn : 1;
1417 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1418 for (int j = start; j < end; j++) {
1419 Node* t1 = s1->in(j);
1420 Node* t2 = s2->in(j);
1421 if (!in_bb(t1) || !in_bb(t2))
1422 continue;
1423 if (stmts_can_pack(t1, t2, align)) {
1424 if (est_savings(t1, t2) >= 0) {
1425 Node_List* pair = new Node_List();
1426 pair->push(t1);
1427 pair->push(t2);
1428 _packset.append(pair);
1429 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1430 set_alignment(t1, t2, align);
1431 changed = true;
1432 }
1433 }
1434 }
1435 return changed;
1436 }
1437
1438 //------------------------------follow_def_uses---------------------------
1439 // Extend the packset by visiting uses of nodes in pack p
1440 bool SuperWord::follow_def_uses(Node_List* p) {
1441 bool changed = false;
1442 Node* s1 = p->at(0);
1443 Node* s2 = p->at(1);
1444 assert(p->size() == 2, "just checking");
1445 assert(s1->req() == s2->req(), "just checking");
1446 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1447
1448 if (s1->is_Store()) return false;
1449
1450 int align = alignment(s1);
1451 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1452 int savings = -1;
1453 int num_s1_uses = 0;
1454 Node* u1 = NULL;
1455 Node* u2 = NULL;
1456 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1457 Node* t1 = s1->fast_out(i);
1458 num_s1_uses++;
1459 if (!in_bb(t1)) continue;
1460 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1461 Node* t2 = s2->fast_out(j);
1462 if (!in_bb(t2)) continue;
1463 if (t2->Opcode() == Op_AddI && t2 == _lp->as_CountedLoop()->incr()) continue; // don't mess with the iv
1464 if (!opnd_positions_match(s1, t1, s2, t2))
1465 continue;
1466 if (stmts_can_pack(t1, t2, align)) {
1467 int my_savings = est_savings(t1, t2);
1468 if (my_savings > savings) {
1469 savings = my_savings;
1470 u1 = t1;
1471 u2 = t2;
1472 }
1473 }
1474 }
1475 }
1476 if (num_s1_uses > 1) {
1477 _race_possible = true;
1478 }
1479 if (savings >= 0) {
1480 Node_List* pair = new Node_List();
1481 pair->push(u1);
1482 pair->push(u2);
1483 _packset.append(pair);
1484 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1485 set_alignment(u1, u2, align);
1486 changed = true;
1487 }
1488 return changed;
1489 }
1490
1491 //------------------------------order_def_uses---------------------------
1492 // For extended packsets, ordinally arrange uses packset by major component
1493 void SuperWord::order_def_uses(Node_List* p) {
1494 Node* s1 = p->at(0);
1495
1496 if (s1->is_Store()) return;
1497
1498 // reductions are always managed beforehand
1499 if (s1->is_reduction()) return;
1500
1501 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1502 Node* t1 = s1->fast_out(i);
1503
1504 // Only allow operand swap on commuting operations
1505 if (!t1->is_Add() && !t1->is_Mul() && !VectorNode::is_muladds2i(t1)) {
1506 break;
1507 }
1508
1509 // Now find t1's packset
1510 Node_List* p2 = NULL;
1511 for (int j = 0; j < _packset.length(); j++) {
1512 p2 = _packset.at(j);
1513 Node* first = p2->at(0);
1514 if (t1 == first) {
1515 break;
1516 }
1517 p2 = NULL;
1518 }
1519 // Arrange all sub components by the major component
1520 if (p2 != NULL) {
1521 for (uint j = 1; j < p->size(); j++) {
1522 Node* d1 = p->at(j);
1523 Node* u1 = p2->at(j);
1524 opnd_positions_match(s1, t1, d1, u1);
1525 }
1526 }
1527 }
1528 }
1529
1530 //---------------------------opnd_positions_match-------------------------
1531 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1532 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1533 // check reductions to see if they are marshalled to represent the reduction
1534 // operator in a specified opnd
1535 if (u1->is_reduction() && u2->is_reduction()) {
1536 // ensure reductions have phis and reduction definitions feeding the 1st operand
1537 Node* first = u1->in(2);
1538 if (first->is_Phi() || first->is_reduction()) {
1539 u1->swap_edges(1, 2);
1540 }
1541 // ensure reductions have phis and reduction definitions feeding the 1st operand
1542 first = u2->in(2);
1543 if (first->is_Phi() || first->is_reduction()) {
1544 u2->swap_edges(1, 2);
1545 }
1546 return true;
1547 }
1548
1549 uint ct = u1->req();
1550 if (ct != u2->req()) return false;
1551 uint i1 = 0;
1552 uint i2 = 0;
1553 do {
1554 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1555 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1556 if (i1 != i2) {
1557 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1558 // Further analysis relies on operands position matching.
1559 u2->swap_edges(i1, i2);
1560 } else if (VectorNode::is_muladds2i(u2) && u1 != u2) {
1561 if (i1 == 5 - i2) { // ((i1 == 3 && i2 == 2) || (i1 == 2 && i2 == 3) || (i1 == 1 && i2 == 4) || (i1 == 4 && i2 == 1))
1562 u2->swap_edges(1, 2);
1563 u2->swap_edges(3, 4);
1564 }
1565 if (i1 == 3 - i2 || i1 == 7 - i2) { // ((i1 == 1 && i2 == 2) || (i1 == 2 && i2 == 1) || (i1 == 3 && i2 == 4) || (i1 == 4 && i2 == 3))
1566 u2->swap_edges(2, 3);
1567 u2->swap_edges(1, 4);
1568 }
1569 return false; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
1570 } else {
1571 return false;
1572 }
1573 } else if (i1 == i2 && VectorNode::is_muladds2i(u2) && u1 != u2) {
1574 u2->swap_edges(1, 3);
1575 u2->swap_edges(2, 4);
1576 return false; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
1577 }
1578 } while (i1 < ct);
1579 return true;
1580 }
1581
1582 //------------------------------est_savings---------------------------
1583 // Estimate the savings from executing s1 and s2 as a pack
1584 int SuperWord::est_savings(Node* s1, Node* s2) {
1585 int save_in = 2 - 1; // 2 operations per instruction in packed form
1586
1587 // inputs
1588 for (uint i = 1; i < s1->req(); i++) {
1589 Node* x1 = s1->in(i);
1590 Node* x2 = s2->in(i);
1591 if (x1 != x2) {
1592 if (are_adjacent_refs(x1, x2)) {
1593 save_in += adjacent_profit(x1, x2);
1594 } else if (!in_packset(x1, x2)) {
1595 save_in -= pack_cost(2);
1596 } else {
1597 save_in += unpack_cost(2);
1598 }
1599 }
1600 }
1601
1602 // uses of result
1603 uint ct = 0;
1604 int save_use = 0;
1605 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1606 Node* s1_use = s1->fast_out(i);
1607 for (int j = 0; j < _packset.length(); j++) {
1608 Node_List* p = _packset.at(j);
1609 if (p->at(0) == s1_use) {
1610 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1611 Node* s2_use = s2->fast_out(k);
1612 if (p->at(p->size()-1) == s2_use) {
1613 ct++;
1614 if (are_adjacent_refs(s1_use, s2_use)) {
1615 save_use += adjacent_profit(s1_use, s2_use);
1616 }
1617 }
1618 }
1619 }
1620 }
1621 }
1622
1623 if (ct < s1->outcnt()) save_use += unpack_cost(1);
1624 if (ct < s2->outcnt()) save_use += unpack_cost(1);
1625
1626 return MAX2(save_in, save_use);
1627 }
1628
1629 //------------------------------costs---------------------------
1630 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1631 int SuperWord::pack_cost(int ct) { return ct; }
1632 int SuperWord::unpack_cost(int ct) { return ct; }
1633
1634 //------------------------------combine_packs---------------------------
1635 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1636 void SuperWord::combine_packs() {
1637 bool changed = true;
1638 // Combine packs regardless max vector size.
1639 while (changed) {
1640 changed = false;
1641 for (int i = 0; i < _packset.length(); i++) {
1642 Node_List* p1 = _packset.at(i);
1643 if (p1 == NULL) continue;
1644 // Because of sorting we can start at i + 1
1645 for (int j = i + 1; j < _packset.length(); j++) {
1646 Node_List* p2 = _packset.at(j);
1647 if (p2 == NULL) continue;
1648 if (i == j) continue;
1649 if (p1->at(p1->size()-1) == p2->at(0)) {
1650 for (uint k = 1; k < p2->size(); k++) {
1651 p1->push(p2->at(k));
1652 }
1653 _packset.at_put(j, NULL);
1654 changed = true;
1655 }
1656 }
1657 }
1658 }
1659
1660 // Split packs which have size greater then max vector size.
1661 for (int i = 0; i < _packset.length(); i++) {
1662 Node_List* p1 = _packset.at(i);
1663 if (p1 != NULL) {
1664 BasicType bt = velt_basic_type(p1->at(0));
1665 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1666 assert(is_power_of_2(max_vlen), "sanity");
1667 uint psize = p1->size();
1668 if (!is_power_of_2(psize)) {
1669 // Skip pack which can't be vector.
1670 // case1: for(...) { a[i] = i; } elements values are different (i+x)
1671 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
1672 _packset.at_put(i, NULL);
1673 continue;
1674 }
1675 if (psize > max_vlen) {
1676 Node_List* pack = new Node_List();
1677 for (uint j = 0; j < psize; j++) {
1678 pack->push(p1->at(j));
1679 if (pack->size() >= max_vlen) {
1680 assert(is_power_of_2(pack->size()), "sanity");
1681 _packset.append(pack);
1682 pack = new Node_List();
1683 }
1684 }
1685 _packset.at_put(i, NULL);
1686 }
1687 }
1688 }
1689
1690 // Compress list.
1691 for (int i = _packset.length() - 1; i >= 0; i--) {
1692 Node_List* p1 = _packset.at(i);
1693 if (p1 == NULL) {
1694 _packset.remove_at(i);
1695 }
1696 }
1697
1698 if (TraceSuperWord) {
1699 tty->print_cr("\nAfter combine_packs");
1700 print_packset();
1701 }
1702 }
1703
1704 //-----------------------------construct_my_pack_map--------------------------
1705 // Construct the map from nodes to packs. Only valid after the
1706 // point where a node is only in one pack (after combine_packs).
1707 void SuperWord::construct_my_pack_map() {
1708 Node_List* rslt = NULL;
1709 for (int i = 0; i < _packset.length(); i++) {
1710 Node_List* p = _packset.at(i);
1711 for (uint j = 0; j < p->size(); j++) {
1712 Node* s = p->at(j);
1713 assert(my_pack(s) == NULL, "only in one pack");
1714 set_my_pack(s, p);
1715 }
1716 }
1717 }
1718
1719 //------------------------------filter_packs---------------------------
1720 // Remove packs that are not implemented or not profitable.
1721 void SuperWord::filter_packs() {
1722 // Remove packs that are not implemented
1723 for (int i = _packset.length() - 1; i >= 0; i--) {
1724 Node_List* pk = _packset.at(i);
1725 bool impl = implemented(pk);
1726 if (!impl) {
1727 #ifndef PRODUCT
1728 if (TraceSuperWord && Verbose) {
1729 tty->print_cr("Unimplemented");
1730 pk->at(0)->dump();
1731 }
1732 #endif
1733 remove_pack_at(i);
1734 }
1735 Node *n = pk->at(0);
1736 if (n->is_reduction()) {
1737 _num_reductions++;
1738 } else {
1739 _num_work_vecs++;
1740 }
1741 }
1742
1743 // Remove packs that are not profitable
1744 bool changed;
1745 do {
1746 changed = false;
1747 for (int i = _packset.length() - 1; i >= 0; i--) {
1748 Node_List* pk = _packset.at(i);
1749 bool prof = profitable(pk);
1750 if (!prof) {
1751 #ifndef PRODUCT
1752 if (TraceSuperWord && Verbose) {
1753 tty->print_cr("Unprofitable");
1754 pk->at(0)->dump();
1755 }
1756 #endif
1757 remove_pack_at(i);
1758 changed = true;
1759 }
1760 }
1761 } while (changed);
1762
1763 #ifndef PRODUCT
1764 if (TraceSuperWord) {
1765 tty->print_cr("\nAfter filter_packs");
1766 print_packset();
1767 tty->cr();
1768 }
1769 #endif
1770 }
1771
1772 //------------------------------merge_packs_to_cmovd---------------------------
1773 // Merge CMoveD into new vector-nodes
1774 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1775 //
1776 // SubD ConD
1777 // / | /
1778 // / | / /
1779 // / | / /
1780 // / | / /
1781 // / / /
1782 // / / | /
1783 // v / | /
1784 // CmpD | /
1785 // | | /
1786 // v | /
1787 // Bool | /
1788 // \ | /
1789 // \ | /
1790 // \ | /
1791 // \ | /
1792 // \ v /
1793 // CMoveD
1794 //
1795
1796 void SuperWord::merge_packs_to_cmovd() {
1797 for (int i = _packset.length() - 1; i >= 0; i--) {
1798 _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1799 }
1800 #ifndef PRODUCT
1801 if (TraceSuperWord) {
1802 tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1803 print_packset();
1804 tty->cr();
1805 }
1806 #endif
1807 }
1808
1809 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1810 Node* use = NULL;
1811 if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1812 return NULL;
1813 }
1814 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1815 use = def->fast_out(j);
1816 if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1817 return NULL;
1818 }
1819 }
1820 return use;
1821 }
1822
1823 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1824 Node* use = NULL;
1825 if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1826 return NULL;
1827 }
1828 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1829 use = def->fast_out(j);
1830 if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1831 return NULL;
1832 }
1833 }
1834 return use;
1835 }
1836
1837 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1838 Node *cmovd = cmovd_pk->at(0);
1839 if (!cmovd->is_CMove()) {
1840 return NULL;
1841 }
1842 if (cmovd->Opcode() != Op_CMoveF && cmovd->Opcode() != Op_CMoveD) {
1843 return NULL;
1844 }
1845 if (pack(cmovd) != NULL) { // already in the cmov pack
1846 return NULL;
1847 }
1848 if (cmovd->in(0) != NULL) {
1849 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1850 return NULL;
1851 }
1852
1853 Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1854 if (!bol->is_Bool()
1855 || bol->outcnt() != 1
1856 || !_sw->same_generation(bol, cmovd)
1857 || bol->in(0) != NULL // BoolNode has control flow!!
1858 || _sw->my_pack(bol) == NULL) {
1859 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();})
1860 return NULL;
1861 }
1862 Node_List* bool_pk = _sw->my_pack(bol);
1863 if (bool_pk->size() != cmovd_pk->size() ) {
1864 return NULL;
1865 }
1866
1867 Node* cmpd = bol->in(1);
1868 if (!cmpd->is_Cmp()
1869 || cmpd->outcnt() != 1
1870 || !_sw->same_generation(cmpd, cmovd)
1871 || cmpd->in(0) != NULL // CmpDNode has control flow!!
1872 || _sw->my_pack(cmpd) == NULL) {
1873 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();})
1874 return NULL;
1875 }
1876 Node_List* cmpd_pk = _sw->my_pack(cmpd);
1877 if (cmpd_pk->size() != cmovd_pk->size() ) {
1878 return NULL;
1879 }
1880
1881 if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1882 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();})
1883 return NULL;
1884 }
1885
1886 Node_List* new_cmpd_pk = new Node_List();
1887 uint sz = cmovd_pk->size() - 1;
1888 for (uint i = 0; i <= sz; ++i) {
1889 Node* cmov = cmovd_pk->at(i);
1890 Node* bol = bool_pk->at(i);
1891 Node* cmp = cmpd_pk->at(i);
1892
1893 new_cmpd_pk->insert(i, cmov);
1894
1895 map(cmov, new_cmpd_pk);
1896 map(bol, new_cmpd_pk);
1897 map(cmp, new_cmpd_pk);
1898
1899 _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1900 }
1901 _sw->_packset.remove(cmovd_pk);
1902 _sw->_packset.remove(bool_pk);
1903 _sw->_packset.remove(cmpd_pk);
1904 _sw->_packset.append(new_cmpd_pk);
1905 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1906 return new_cmpd_pk;
1907 }
1908
1909 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1910 Node* cmpd0 = cmpd_pk->at(0);
1911 assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1912 assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1913 assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1914 Node* in1 = cmpd0->in(1);
1915 Node* in2 = cmpd0->in(2);
1916 Node_List* in1_pk = _sw->my_pack(in1);
1917 Node_List* in2_pk = _sw->my_pack(in2);
1918
1919 if ( (in1_pk != NULL && in1_pk->size() != cmpd_pk->size())
1920 || (in2_pk != NULL && in2_pk->size() != cmpd_pk->size()) ) {
1921 return false;
1922 }
1923
1924 // test if "all" in1 are in the same pack or the same node
1925 if (in1_pk == NULL) {
1926 for (uint j = 1; j < cmpd_pk->size(); j++) {
1927 if (cmpd_pk->at(j)->in(1) != in1) {
1928 return false;
1929 }
1930 }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1931 }
1932 // test if "all" in2 are in the same pack or the same node
1933 if (in2_pk == NULL) {
1934 for (uint j = 1; j < cmpd_pk->size(); j++) {
1935 if (cmpd_pk->at(j)->in(2) != in2) {
1936 return false;
1937 }
1938 }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1939 }
1940 //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1941 int cmovd_ind1, cmovd_ind2;
1942 if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1943 && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1944 cmovd_ind1 = CMoveNode::IfFalse;
1945 cmovd_ind2 = CMoveNode::IfTrue;
1946 } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1947 && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1948 cmovd_ind2 = CMoveNode::IfFalse;
1949 cmovd_ind1 = CMoveNode::IfTrue;
1950 }
1951 else {
1952 return false;
1953 }
1954
1955 for (uint j = 1; j < cmpd_pk->size(); j++) {
1956 if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1957 || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1958 return false;
1959 }//if
1960 }
1961 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(); })
1962 return true;
1963 }
1964
1965 //------------------------------implemented---------------------------
1966 // Can code be generated for pack p?
1967 bool SuperWord::implemented(Node_List* p) {
1968 bool retValue = false;
1969 Node* p0 = p->at(0);
1970 if (p0 != NULL) {
1971 int opc = p0->Opcode();
1972 uint size = p->size();
1973 if (p0->is_reduction()) {
1974 const Type *arith_type = p0->bottom_type();
1975 // Length 2 reductions of INT/LONG do not offer performance benefits
1976 if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1977 retValue = false;
1978 } else {
1979 retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1980 }
1981 } else {
1982 retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1983 }
1984 if (!retValue) {
1985 if (is_cmov_pack(p)) {
1986 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1987 return true;
1988 }
1989 }
1990 }
1991 return retValue;
1992 }
1993
1994 bool SuperWord::is_cmov_pack(Node_List* p) {
1995 return _cmovev_kit.pack(p->at(0)) != NULL;
1996 }
1997 //------------------------------same_inputs--------------------------
1998 // For pack p, are all idx operands the same?
1999 bool SuperWord::same_inputs(Node_List* p, int idx) {
2000 Node* p0 = p->at(0);
2001 uint vlen = p->size();
2002 Node* p0_def = p0->in(idx);
2003 for (uint i = 1; i < vlen; i++) {
2004 Node* pi = p->at(i);
2005 Node* pi_def = pi->in(idx);
2006 if (p0_def != pi_def) {
2007 return false;
2008 }
2009 }
2010 return true;
2011 }
2012
2013 //------------------------------profitable---------------------------
2014 // For pack p, are all operands and all uses (with in the block) vector?
2015 bool SuperWord::profitable(Node_List* p) {
2016 Node* p0 = p->at(0);
2017 uint start, end;
2018 VectorNode::vector_operands(p0, &start, &end);
2019
2020 // Return false if some inputs are not vectors or vectors with different
2021 // size or alignment.
2022 // Also, for now, return false if not scalar promotion case when inputs are
2023 // the same. Later, implement PackNode and allow differing, non-vector inputs
2024 // (maybe just the ones from outside the block.)
2025 for (uint i = start; i < end; i++) {
2026 if (!is_vector_use(p0, i)) {
2027 return false;
2028 }
2029 }
2030 // Check if reductions are connected
2031 if (p0->is_reduction()) {
2032 Node* second_in = p0->in(2);
2033 Node_List* second_pk = my_pack(second_in);
2034 if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
2035 // Remove reduction flag if no parent pack or if not enough work
2036 // to cover reduction expansion overhead
2037 p0->remove_flag(Node::Flag_is_reduction);
2038 return false;
2039 } else if (second_pk->size() != p->size()) {
2040 return false;
2041 }
2042 }
2043 if (VectorNode::is_shift(p0)) {
2044 // For now, return false if shift count is vector or not scalar promotion
2045 // case (different shift counts) because it is not supported yet.
2046 Node* cnt = p0->in(2);
2047 Node_List* cnt_pk = my_pack(cnt);
2048 if (cnt_pk != NULL)
2049 return false;
2050 if (!same_inputs(p, 2))
2051 return false;
2052 }
2053 if (!p0->is_Store()) {
2054 // For now, return false if not all uses are vector.
2055 // Later, implement ExtractNode and allow non-vector uses (maybe
2056 // just the ones outside the block.)
2057 for (uint i = 0; i < p->size(); i++) {
2058 Node* def = p->at(i);
2059 if (is_cmov_pack_internal_node(p, def)) {
2060 continue;
2061 }
2062 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2063 Node* use = def->fast_out(j);
2064 for (uint k = 0; k < use->req(); k++) {
2065 Node* n = use->in(k);
2066 if (def == n) {
2067 // Reductions should only have a Phi use at the loop head or a non-phi use
2068 // outside of the loop if it is the last element of the pack (e.g. SafePoint).
2069 if (def->is_reduction() &&
2070 ((use->is_Phi() && use->in(0) == _lpt->_head) ||
2071 (!_lpt->is_member(_phase->get_loop(_phase->ctrl_or_self(use))) && i == p->size()-1))) {
2072 continue;
2073 }
2074 if (!is_vector_use(use, k)) {
2075 return false;
2076 }
2077 }
2078 }
2079 }
2080 }
2081 }
2082 return true;
2083 }
2084
2085 //------------------------------schedule---------------------------
2086 // Adjust the memory graph for the packed operations
2087 void SuperWord::schedule() {
2088
2089 // Co-locate in the memory graph the members of each memory pack
2090 for (int i = 0; i < _packset.length(); i++) {
2091 co_locate_pack(_packset.at(i));
2092 }
2093 }
2094
2095 //-------------------------------remove_and_insert-------------------
2096 // Remove "current" from its current position in the memory graph and insert
2097 // it after the appropriate insertion point (lip or uip).
2098 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
2099 Node *uip, Unique_Node_List &sched_before) {
2100 Node* my_mem = current->in(MemNode::Memory);
2101 bool sched_up = sched_before.member(current);
2102
2103 // remove current_store from its current position in the memmory graph
2104 for (DUIterator i = current->outs(); current->has_out(i); i++) {
2105 Node* use = current->out(i);
2106 if (use->is_Mem()) {
2107 assert(use->in(MemNode::Memory) == current, "must be");
2108 if (use == prev) { // connect prev to my_mem
2109 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
2110 --i; //deleted this edge; rescan position
2111 } else if (sched_before.member(use)) {
2112 if (!sched_up) { // Will be moved together with current
2113 _igvn.replace_input_of(use, MemNode::Memory, uip);
2114 --i; //deleted this edge; rescan position
2115 }
2116 } else {
2117 if (sched_up) { // Will be moved together with current
2118 _igvn.replace_input_of(use, MemNode::Memory, lip);
2119 --i; //deleted this edge; rescan position
2120 }
2121 }
2122 }
2123 }
2124
2125 Node *insert_pt = sched_up ? uip : lip;
2126
2127 // all uses of insert_pt's memory state should use current's instead
2128 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
2129 Node* use = insert_pt->out(i);
2130 if (use->is_Mem()) {
2131 assert(use->in(MemNode::Memory) == insert_pt, "must be");
2132 _igvn.replace_input_of(use, MemNode::Memory, current);
2133 --i; //deleted this edge; rescan position
2134 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
2135 uint pos; //lip (lower insert point) must be the last one in the memory slice
2136 for (pos=1; pos < use->req(); pos++) {
2137 if (use->in(pos) == insert_pt) break;
2138 }
2139 _igvn.replace_input_of(use, pos, current);
2140 --i;
2141 }
2142 }
2143
2144 //connect current to insert_pt
2145 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
2146 }
2147
2148 //------------------------------co_locate_pack----------------------------------
2149 // To schedule a store pack, we need to move any sandwiched memory ops either before
2150 // or after the pack, based upon dependence information:
2151 // (1) If any store in the pack depends on the sandwiched memory op, the
2152 // sandwiched memory op must be scheduled BEFORE the pack;
2153 // (2) If a sandwiched memory op depends on any store in the pack, the
2154 // sandwiched memory op must be scheduled AFTER the pack;
2155 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
2156 // memory op (say memB), memB must be scheduled before memA. So, if memA is
2157 // scheduled before the pack, memB must also be scheduled before the pack;
2158 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
2159 // schedule this store AFTER the pack
2160 // (5) We know there is no dependence cycle, so there in no other case;
2161 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
2162 //
2163 // To schedule a load pack, we use the memory state of either the first or the last load in
2164 // the pack, based on the dependence constraint.
2165 void SuperWord::co_locate_pack(Node_List* pk) {
2166 if (pk->at(0)->is_Store()) {
2167 MemNode* first = executed_first(pk)->as_Mem();
2168 MemNode* last = executed_last(pk)->as_Mem();
2169 Unique_Node_List schedule_before_pack;
2170 Unique_Node_List memops;
2171
2172 MemNode* current = last->in(MemNode::Memory)->as_Mem();
2173 MemNode* previous = last;
2174 while (true) {
2175 assert(in_bb(current), "stay in block");
2176 memops.push(previous);
2177 for (DUIterator i = current->outs(); current->has_out(i); i++) {
2178 Node* use = current->out(i);
2179 if (use->is_Mem() && use != previous)
2180 memops.push(use);
2181 }
2182 if (current == first) break;
2183 previous = current;
2184 current = current->in(MemNode::Memory)->as_Mem();
2185 }
2186
2187 // determine which memory operations should be scheduled before the pack
2188 for (uint i = 1; i < memops.size(); i++) {
2189 Node *s1 = memops.at(i);
2190 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
2191 for (uint j = 0; j< i; j++) {
2192 Node *s2 = memops.at(j);
2193 if (!independent(s1, s2)) {
2194 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
2195 schedule_before_pack.push(s1); // s1 must be scheduled before
2196 Node_List* mem_pk = my_pack(s1);
2197 if (mem_pk != NULL) {
2198 for (uint ii = 0; ii < mem_pk->size(); ii++) {
2199 Node* s = mem_pk->at(ii); // follow partner
2200 if (memops.member(s) && !schedule_before_pack.member(s))
2201 schedule_before_pack.push(s);
2202 }
2203 }
2204 break;
2205 }
2206 }
2207 }
2208 }
2209 }
2210
2211 Node* upper_insert_pt = first->in(MemNode::Memory);
2212 // Following code moves loads connected to upper_insert_pt below aliased stores.
2213 // Collect such loads here and reconnect them back to upper_insert_pt later.
2214 memops.clear();
2215 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
2216 Node* use = upper_insert_pt->out(i);
2217 if (use->is_Mem() && !use->is_Store()) {
2218 memops.push(use);
2219 }
2220 }
2221
2222 MemNode* lower_insert_pt = last;
2223 previous = last; //previous store in pk
2224 current = last->in(MemNode::Memory)->as_Mem();
2225
2226 // start scheduling from "last" to "first"
2227 while (true) {
2228 assert(in_bb(current), "stay in block");
2229 assert(in_pack(previous, pk), "previous stays in pack");
2230 Node* my_mem = current->in(MemNode::Memory);
2231
2232 if (in_pack(current, pk)) {
2233 // Forward users of my memory state (except "previous) to my input memory state
2234 for (DUIterator i = current->outs(); current->has_out(i); i++) {
2235 Node* use = current->out(i);
2236 if (use->is_Mem() && use != previous) {
2237 assert(use->in(MemNode::Memory) == current, "must be");
2238 if (schedule_before_pack.member(use)) {
2239 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
2240 } else {
2241 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
2242 }
2243 --i; // deleted this edge; rescan position
2244 }
2245 }
2246 previous = current;
2247 } else { // !in_pack(current, pk) ==> a sandwiched store
2248 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
2249 }
2250
2251 if (current == first) break;
2252 current = my_mem->as_Mem();
2253 } // end while
2254
2255 // Reconnect loads back to upper_insert_pt.
2256 for (uint i = 0; i < memops.size(); i++) {
2257 Node *ld = memops.at(i);
2258 if (ld->in(MemNode::Memory) != upper_insert_pt) {
2259 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
2260 }
2261 }
2262 } else if (pk->at(0)->is_Load()) { // Load pack
2263 // All loads in the pack should have the same memory state. By default,
2264 // we use the memory state of the last load. However, if any load could
2265 // not be moved down due to the dependence constraint, we use the memory
2266 // state of the first load.
2267 Node* mem_input = pick_mem_state(pk);
2268 _igvn.hash_delete(mem_input);
2269 // Give each load the same memory state
2270 for (uint i = 0; i < pk->size(); i++) {
2271 LoadNode* ld = pk->at(i)->as_Load();
2272 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
2273 }
2274 }
2275 }
2276
2277 // Finds the first and last memory state and then picks either of them by checking dependence constraints.
2278 // If a store is dependent on an earlier load then we need to pick the memory state of the first load and cannot
2279 // pick the memory state of the last load.
2280 Node* SuperWord::pick_mem_state(Node_List* pk) {
2281 Node* first_mem = find_first_mem_state(pk);
2282 Node* last_mem = find_last_mem_state(pk, first_mem);
2283
2284 for (uint i = 0; i < pk->size(); i++) {
2285 Node* ld = pk->at(i);
2286 for (Node* current = last_mem; current != ld->in(MemNode::Memory); current = current->in(MemNode::Memory)) {
2287 assert(current->is_Mem() && in_bb(current), "unexpected memory");
2288 assert(current != first_mem, "corrupted memory graph");
2289 if (!independent(current, ld)) {
2290 #ifdef ASSERT
2291 // Added assertion code since no case has been observed that should pick the first memory state.
2292 // Remove the assertion code whenever we find a (valid) case that really needs the first memory state.
2293 pk->dump();
2294 first_mem->dump();
2295 last_mem->dump();
2296 current->dump();
2297 ld->dump();
2298 ld->in(MemNode::Memory)->dump();
2299 assert(false, "never observed that first memory should be picked");
2300 #endif
2301 return first_mem; // A later store depends on this load, pick memory state of first load
2302 }
2303 }
2304 }
2305 return last_mem;
2306 }
2307
2308 // Walk the memory graph from the current first load until the
2309 // start of the loop and check if nodes on the way are memory
2310 // edges of loads in the pack. The last one we encounter is the
2311 // first load.
2312 Node* SuperWord::find_first_mem_state(Node_List* pk) {
2313 Node* first_mem = pk->at(0)->in(MemNode::Memory);
2314 for (Node* current = first_mem; in_bb(current); current = current->is_Phi() ? current->in(LoopNode::EntryControl) : current->in(MemNode::Memory)) {
2315 assert(current->is_Mem() || (current->is_Phi() && current->in(0) == bb()), "unexpected memory");
2316 for (uint i = 1; i < pk->size(); i++) {
2317 Node* ld = pk->at(i);
2318 if (ld->in(MemNode::Memory) == current) {
2319 first_mem = current;
2320 break;
2321 }
2322 }
2323 }
2324 return first_mem;
2325 }
2326
2327 // Find the last load by going over the pack again and walking
2328 // the memory graph from the loads of the pack to the memory of
2329 // the first load. If we encounter the memory of the current last
2330 // load, then we started from further down in the memory graph and
2331 // the load we started from is the last load.
2332 Node* SuperWord::find_last_mem_state(Node_List* pk, Node* first_mem) {
2333 Node* last_mem = pk->at(0)->in(MemNode::Memory);
2334 for (uint i = 0; i < pk->size(); i++) {
2335 Node* ld = pk->at(i);
2336 for (Node* current = ld->in(MemNode::Memory); current != first_mem; current = current->in(MemNode::Memory)) {
2337 assert(current->is_Mem() && in_bb(current), "unexpected memory");
2338 if (current->in(MemNode::Memory) == last_mem) {
2339 last_mem = ld->in(MemNode::Memory);
2340 }
2341 }
2342 }
2343 return last_mem;
2344 }
2345
2346 #ifndef PRODUCT
2347 void SuperWord::print_loop(bool whole) {
2348 Node_Stack stack(_arena, _phase->C->unique() >> 2);
2349 Node_List rpo_list;
2350 VectorSet visited(_arena);
2351 visited.set(lpt()->_head->_idx);
2352 _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2353 _phase->dump(lpt(), rpo_list.size(), rpo_list );
2354 if(whole) {
2355 tty->print_cr("\n Whole loop tree");
2356 _phase->dump();
2357 tty->print_cr(" End of whole loop tree\n");
2358 }
2359 }
2360 #endif
2361
2362 //------------------------------output---------------------------
2363 // Convert packs into vector node operations
2364 void SuperWord::output() {
2365 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2366 Compile* C = _phase->C;
2367 if (_packset.length() == 0) {
2368 if (cl->is_main_loop()) {
2369 // Instigate more unrolling for optimization when vectorization fails.
2370 C->set_major_progress();
2371 cl->set_notpassed_slp();
2372 cl->mark_do_unroll_only();
2373 }
2374 return;
2375 }
2376
2377 #ifndef PRODUCT
2378 if (TraceLoopOpts) {
2379 tty->print("SuperWord::output ");
2380 lpt()->dump_head();
2381 }
2382 #endif
2383
2384 if (cl->is_main_loop()) {
2385 // MUST ENSURE main loop's initial value is properly aligned:
2386 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2387
2388 align_initial_loop_index(align_to_ref());
2389
2390 // Insert extract (unpack) operations for scalar uses
2391 for (int i = 0; i < _packset.length(); i++) {
2392 insert_extracts(_packset.at(i));
2393 }
2394 }
2395
2396 uint max_vlen_in_bytes = 0;
2397 uint max_vlen = 0;
2398 bool can_process_post_loop = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
2399
2400 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2401
2402 CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2403
2404 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2405
2406 if (do_reserve_copy() && !make_reversable.has_reserved()) {
2407 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2408 return;
2409 }
2410
2411 for (int i = 0; i < _block.length(); i++) {
2412 Node* n = _block.at(i);
2413 Node_List* p = my_pack(n);
2414 if (p && n == executed_last(p)) {
2415 uint vlen = p->size();
2416 uint vlen_in_bytes = 0;
2417 Node* vn = NULL;
2418 Node* low_adr = p->at(0);
2419 Node* first = executed_first(p);
2420 if (can_process_post_loop) {
2421 // override vlen with the main loops vector length
2422 vlen = cl->slp_max_unroll();
2423 }
2424 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);})
2425 int opc = n->Opcode();
2426 if (n->is_Load()) {
2427 Node* ctl = n->in(MemNode::Control);
2428 Node* mem = first->in(MemNode::Memory);
2429 SWPointer p1(n->as_Mem(), this, NULL, false);
2430 // Identify the memory dependency for the new loadVector node by
2431 // walking up through memory chain.
2432 // This is done to give flexibility to the new loadVector node so that
2433 // it can move above independent storeVector nodes.
2434 while (mem->is_StoreVector()) {
2435 SWPointer p2(mem->as_Mem(), this, NULL, false);
2436 int cmp = p1.cmp(p2);
2437 if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2438 mem = mem->in(MemNode::Memory);
2439 } else {
2440 break; // dependent memory
2441 }
2442 }
2443 Node* adr = low_adr->in(MemNode::Address);
2444 const TypePtr* atyp = n->adr_type();
2445 vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2446 vlen_in_bytes = vn->as_LoadVector()->memory_size();
2447 } else if (n->is_Store()) {
2448 // Promote value to be stored to vector
2449 Node* val = vector_opd(p, MemNode::ValueIn);
2450 if (val == NULL) {
2451 if (do_reserve_copy()) {
2452 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2453 return; //and reverse to backup IG
2454 }
2455 ShouldNotReachHere();
2456 }
2457
2458 Node* ctl = n->in(MemNode::Control);
2459 Node* mem = first->in(MemNode::Memory);
2460 Node* adr = low_adr->in(MemNode::Address);
2461 const TypePtr* atyp = n->adr_type();
2462 vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2463 vlen_in_bytes = vn->as_StoreVector()->memory_size();
2464 } else if (VectorNode::is_roundopD(n)) {
2465 Node* in1 = vector_opd(p, 1);
2466 Node* in2 = low_adr->in(2);
2467 assert(in2->is_Con(), "Constant rounding mode expected.");
2468 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2469 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2470 } else if (VectorNode::is_muladds2i(n)) {
2471 assert(n->req() == 5u, "MulAddS2I should have 4 operands.");
2472 Node* in1 = vector_opd(p, 1);
2473 Node* in2 = vector_opd(p, 2);
2474 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2475 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2476 } else if (n->req() == 3 && !is_cmov_pack(p)) {
2477 // Promote operands to vector
2478 Node* in1 = NULL;
2479 bool node_isa_reduction = n->is_reduction();
2480 if (node_isa_reduction) {
2481 // the input to the first reduction operation is retained
2482 in1 = low_adr->in(1);
2483 } else {
2484 in1 = vector_opd(p, 1);
2485 if (in1 == NULL) {
2486 if (do_reserve_copy()) {
2487 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2488 return; //and reverse to backup IG
2489 }
2490 ShouldNotReachHere();
2491 }
2492 }
2493 Node* in2 = vector_opd(p, 2);
2494 if (in2 == NULL) {
2495 if (do_reserve_copy()) {
2496 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2497 return; //and reverse to backup IG
2498 }
2499 ShouldNotReachHere();
2500 }
2501 if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2502 // Move invariant vector input into second position to avoid register spilling.
2503 Node* tmp = in1;
2504 in1 = in2;
2505 in2 = tmp;
2506 }
2507 if (node_isa_reduction) {
2508 const Type *arith_type = n->bottom_type();
2509 vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2510 if (in2->is_Load()) {
2511 vlen_in_bytes = in2->as_LoadVector()->memory_size();
2512 } else {
2513 vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2514 }
2515 } else {
2516 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2517 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2518 }
2519 } else if (opc == Op_SqrtF || opc == Op_SqrtD ||
2520 opc == Op_AbsF || opc == Op_AbsD ||
2521 opc == Op_AbsI || opc == Op_AbsL ||
2522 opc == Op_NegF || opc == Op_NegD ||
2523 opc == Op_PopCountI) {
2524 assert(n->req() == 2, "only one input expected");
2525 Node* in = vector_opd(p, 1);
2526 vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2527 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2528 } else if (is_cmov_pack(p)) {
2529 if (can_process_post_loop) {
2530 // do not refactor of flow in post loop context
2531 return;
2532 }
2533 if (!n->is_CMove()) {
2534 continue;
2535 }
2536 // place here CMoveVDNode
2537 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2538 Node* bol = n->in(CMoveNode::Condition);
2539 if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2540 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();})
2541 bol = bol->in(1); //may be ExtractNode
2542 }
2543
2544 assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2545 if (!bol->is_Bool()) {
2546 if (do_reserve_copy()) {
2547 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2548 return; //and reverse to backup IG
2549 }
2550 ShouldNotReachHere();
2551 }
2552
2553 int cond = (int)bol->as_Bool()->_test._test;
2554 Node* in_cc = _igvn.intcon(cond);
2555 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2556 Node* cc = bol->clone();
2557 cc->set_req(1, in_cc);
2558 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2559
2560 Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2561 if (src1 == NULL) {
2562 if (do_reserve_copy()) {
2563 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2564 return; //and reverse to backup IG
2565 }
2566 ShouldNotReachHere();
2567 }
2568 Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2569 if (src2 == NULL) {
2570 if (do_reserve_copy()) {
2571 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2572 return; //and reverse to backup IG
2573 }
2574 ShouldNotReachHere();
2575 }
2576 BasicType bt = velt_basic_type(n);
2577 const TypeVect* vt = TypeVect::make(bt, vlen);
2578 assert(bt == T_FLOAT || bt == T_DOUBLE, "Only vectorization for FP cmovs is supported");
2579 if (bt == T_FLOAT) {
2580 vn = new CMoveVFNode(cc, src1, src2, vt);
2581 } else {
2582 assert(bt == T_DOUBLE, "Expected double");
2583 vn = new CMoveVDNode(cc, src1, src2, vt);
2584 }
2585 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2586 } else if (opc == Op_FmaD || opc == Op_FmaF) {
2587 // Promote operands to vector
2588 Node* in1 = vector_opd(p, 1);
2589 Node* in2 = vector_opd(p, 2);
2590 Node* in3 = vector_opd(p, 3);
2591 vn = VectorNode::make(opc, in1, in2, in3, vlen, velt_basic_type(n));
2592 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2593 } else {
2594 if (do_reserve_copy()) {
2595 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2596 return; //and reverse to backup IG
2597 }
2598 ShouldNotReachHere();
2599 }
2600
2601 assert(vn != NULL, "sanity");
2602 if (vn == NULL) {
2603 if (do_reserve_copy()){
2604 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2605 return; //and reverse to backup IG
2606 }
2607 ShouldNotReachHere();
2608 }
2609
2610 _block.at_put(i, vn);
2611 _igvn.register_new_node_with_optimizer(vn);
2612 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2613 for (uint j = 0; j < p->size(); j++) {
2614 Node* pm = p->at(j);
2615 _igvn.replace_node(pm, vn);
2616 }
2617 _igvn._worklist.push(vn);
2618
2619 if (can_process_post_loop) {
2620 // first check if the vector size if the maximum vector which we can use on the machine,
2621 // other vector size have reduced values for predicated data mapping.
2622 if (vlen_in_bytes != (uint)MaxVectorSize) {
2623 return;
2624 }
2625 }
2626
2627 if (vlen_in_bytes >= max_vlen_in_bytes && vlen > max_vlen) {
2628 max_vlen = vlen;
2629 max_vlen_in_bytes = vlen_in_bytes;
2630 }
2631 #ifdef ASSERT
2632 if (TraceNewVectors) {
2633 tty->print("new Vector node: ");
2634 vn->dump();
2635 }
2636 #endif
2637 }
2638 }//for (int i = 0; i < _block.length(); i++)
2639
2640 if (max_vlen_in_bytes > C->max_vector_size()) {
2641 C->set_max_vector_size(max_vlen_in_bytes);
2642 }
2643 if (max_vlen_in_bytes > 0) {
2644 cl->mark_loop_vectorized();
2645 }
2646
2647 if (SuperWordLoopUnrollAnalysis) {
2648 if (cl->has_passed_slp()) {
2649 uint slp_max_unroll_factor = cl->slp_max_unroll();
2650 if (slp_max_unroll_factor == max_vlen) {
2651 if (TraceSuperWordLoopUnrollAnalysis) {
2652 tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2653 }
2654
2655 // For atomic unrolled loops which are vector mapped, instigate more unrolling
2656 cl->set_notpassed_slp();
2657 if (cl->is_main_loop()) {
2658 // if vector resources are limited, do not allow additional unrolling, also
2659 // do not unroll more on pure vector loops which were not reduced so that we can
2660 // program the post loop to single iteration execution.
2661 if (FLOATPRESSURE > 8) {
2662 C->set_major_progress();
2663 cl->mark_do_unroll_only();
2664 }
2665 }
2666
2667 if (do_reserve_copy()) {
2668 if (can_process_post_loop) {
2669 // Now create the difference of trip and limit and use it as our mask index.
2670 // Note: We limited the unroll of the vectorized loop so that
2671 // only vlen-1 size iterations can remain to be mask programmed.
2672 Node *incr = cl->incr();
2673 SubINode *index = new SubINode(cl->limit(), cl->init_trip());
2674 _igvn.register_new_node_with_optimizer(index);
2675 SetVectMaskINode *mask = new SetVectMaskINode(_phase->get_ctrl(cl->init_trip()), index);
2676 _igvn.register_new_node_with_optimizer(mask);
2677 // make this a single iteration loop
2678 AddINode *new_incr = new AddINode(incr->in(1), mask);
2679 _igvn.register_new_node_with_optimizer(new_incr);
2680 _phase->set_ctrl(new_incr, _phase->get_ctrl(incr));
2681 _igvn.replace_node(incr, new_incr);
2682 cl->mark_is_multiversioned();
2683 cl->loopexit()->add_flag(Node::Flag_has_vector_mask_set);
2684 }
2685 }
2686 }
2687 }
2688 }
2689
2690 if (do_reserve_copy()) {
2691 make_reversable.use_new();
2692 }
2693 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2694 return;
2695 }
2696
2697 //------------------------------vector_opd---------------------------
2698 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2699 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2700 Node* p0 = p->at(0);
2701 uint vlen = p->size();
2702 Node* opd = p0->in(opd_idx);
2703 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2704
2705 if (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()) {
2706 // override vlen with the main loops vector length
2707 vlen = cl->slp_max_unroll();
2708 }
2709
2710 if (same_inputs(p, opd_idx)) {
2711 if (opd->is_Vector() || opd->is_LoadVector()) {
2712 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2713 if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2714 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2715 return NULL;
2716 }
2717 return opd; // input is matching vector
2718 }
2719 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2720 Compile* C = _phase->C;
2721 Node* cnt = opd;
2722 // Vector instructions do not mask shift count, do it here.
2723 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2724 const TypeInt* t = opd->find_int_type();
2725 if (t != NULL && t->is_con()) {
2726 juint shift = t->get_con();
2727 if (shift > mask) { // Unsigned cmp
2728 cnt = ConNode::make(TypeInt::make(shift & mask));
2729 }
2730 } else {
2731 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2732 cnt = ConNode::make(TypeInt::make(mask));
2733 _igvn.register_new_node_with_optimizer(cnt);
2734 cnt = new AndINode(opd, cnt);
2735 _igvn.register_new_node_with_optimizer(cnt);
2736 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2737 }
2738 assert(opd->bottom_type()->isa_int(), "int type only");
2739 if (!opd->bottom_type()->isa_int()) {
2740 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2741 return NULL;
2742 }
2743 }
2744 // Move shift count into vector register.
2745 cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2746 _igvn.register_new_node_with_optimizer(cnt);
2747 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2748 return cnt;
2749 }
2750 assert(!opd->is_StoreVector(), "such vector is not expected here");
2751 if (opd->is_StoreVector()) {
2752 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2753 return NULL;
2754 }
2755 // Convert scalar input to vector with the same number of elements as
2756 // p0's vector. Use p0's type because size of operand's container in
2757 // vector should match p0's size regardless operand's size.
2758 const Type* p0_t = velt_type(p0);
2759 VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2760
2761 _igvn.register_new_node_with_optimizer(vn);
2762 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2763 #ifdef ASSERT
2764 if (TraceNewVectors) {
2765 tty->print("new Vector node: ");
2766 vn->dump();
2767 }
2768 #endif
2769 return vn;
2770 }
2771
2772 // Insert pack operation
2773 BasicType bt = velt_basic_type(p0);
2774 PackNode* pk = PackNode::make(opd, vlen, bt);
2775 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2776
2777 for (uint i = 1; i < vlen; i++) {
2778 Node* pi = p->at(i);
2779 Node* in = pi->in(opd_idx);
2780 assert(my_pack(in) == NULL, "Should already have been unpacked");
2781 if (my_pack(in) != NULL) {
2782 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2783 return NULL;
2784 }
2785 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2786 pk->add_opd(in);
2787 if (VectorNode::is_muladds2i(pi)) {
2788 Node* in2 = pi->in(opd_idx + 2);
2789 assert(my_pack(in2) == NULL, "Should already have been unpacked");
2790 if (my_pack(in2) != NULL) {
2791 NOT_PRODUCT(if (is_trace_loop_reverse() || TraceLoopOpts) { tty->print_cr("Should already have been unpacked"); })
2792 return NULL;
2793 }
2794 assert(opd_bt == in2->bottom_type()->basic_type(), "all same type");
2795 pk->add_opd(in2);
2796 }
2797 }
2798 _igvn.register_new_node_with_optimizer(pk);
2799 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2800 #ifdef ASSERT
2801 if (TraceNewVectors) {
2802 tty->print("new Vector node: ");
2803 pk->dump();
2804 }
2805 #endif
2806 return pk;
2807 }
2808
2809 //------------------------------insert_extracts---------------------------
2810 // If a use of pack p is not a vector use, then replace the
2811 // use with an extract operation.
2812 void SuperWord::insert_extracts(Node_List* p) {
2813 if (p->at(0)->is_Store()) return;
2814 assert(_n_idx_list.is_empty(), "empty (node,index) list");
2815
2816 // Inspect each use of each pack member. For each use that is
2817 // not a vector use, replace the use with an extract operation.
2818
2819 for (uint i = 0; i < p->size(); i++) {
2820 Node* def = p->at(i);
2821 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2822 Node* use = def->fast_out(j);
2823 for (uint k = 0; k < use->req(); k++) {
2824 Node* n = use->in(k);
2825 if (def == n) {
2826 Node_List* u_pk = my_pack(use);
2827 if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2828 _n_idx_list.push(use, k);
2829 }
2830 }
2831 }
2832 }
2833 }
2834
2835 while (_n_idx_list.is_nonempty()) {
2836 Node* use = _n_idx_list.node();
2837 int idx = _n_idx_list.index();
2838 _n_idx_list.pop();
2839 Node* def = use->in(idx);
2840
2841 if (def->is_reduction()) continue;
2842
2843 // Insert extract operation
2844 _igvn.hash_delete(def);
2845 int def_pos = alignment(def) / data_size(def);
2846
2847 Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2848 _igvn.register_new_node_with_optimizer(ex);
2849 _phase->set_ctrl(ex, _phase->get_ctrl(def));
2850 _igvn.replace_input_of(use, idx, ex);
2851 _igvn._worklist.push(def);
2852
2853 bb_insert_after(ex, bb_idx(def));
2854 set_velt_type(ex, velt_type(def));
2855 }
2856 }
2857
2858 //------------------------------is_vector_use---------------------------
2859 // Is use->in(u_idx) a vector use?
2860 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2861 Node_List* u_pk = my_pack(use);
2862 if (u_pk == NULL) return false;
2863 if (use->is_reduction()) return true;
2864 Node* def = use->in(u_idx);
2865 Node_List* d_pk = my_pack(def);
2866 if (d_pk == NULL) {
2867 // check for scalar promotion
2868 Node* n = u_pk->at(0)->in(u_idx);
2869 for (uint i = 1; i < u_pk->size(); i++) {
2870 if (u_pk->at(i)->in(u_idx) != n) return false;
2871 }
2872 return true;
2873 }
2874 if (VectorNode::is_muladds2i(use)) {
2875 // MulAddS2I takes shorts and produces ints - hence the special checks
2876 // on alignment and size.
2877 if (u_pk->size() * 2 != d_pk->size()) {
2878 return false;
2879 }
2880 for (uint i = 0; i < MIN2(d_pk->size(), u_pk->size()); i++) {
2881 Node* ui = u_pk->at(i);
2882 Node* di = d_pk->at(i);
2883 if (alignment(ui) != alignment(di) * 2) {
2884 return false;
2885 }
2886 }
2887 return true;
2888 }
2889 if (u_pk->size() != d_pk->size())
2890 return false;
2891 for (uint i = 0; i < u_pk->size(); i++) {
2892 Node* ui = u_pk->at(i);
2893 Node* di = d_pk->at(i);
2894 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2895 return false;
2896 }
2897 return true;
2898 }
2899
2900 //------------------------------construct_bb---------------------------
2901 // Construct reverse postorder list of block members
2902 bool SuperWord::construct_bb() {
2903 Node* entry = bb();
2904
2905 assert(_stk.length() == 0, "stk is empty");
2906 assert(_block.length() == 0, "block is empty");
2907 assert(_data_entry.length() == 0, "data_entry is empty");
2908 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2909 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2910
2911 // Find non-control nodes with no inputs from within block,
2912 // create a temporary map from node _idx to bb_idx for use
2913 // by the visited and post_visited sets,
2914 // and count number of nodes in block.
2915 int bb_ct = 0;
2916 for (uint i = 0; i < lpt()->_body.size(); i++) {
2917 Node *n = lpt()->_body.at(i);
2918 set_bb_idx(n, i); // Create a temporary map
2919 if (in_bb(n)) {
2920 if (n->is_LoadStore() || n->is_MergeMem() ||
2921 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2922 // Bailout if the loop has LoadStore, MergeMem or data Proj
2923 // nodes. Superword optimization does not work with them.
2924 return false;
2925 }
2926 bb_ct++;
2927 if (!n->is_CFG()) {
2928 bool found = false;
2929 for (uint j = 0; j < n->req(); j++) {
2930 Node* def = n->in(j);
2931 if (def && in_bb(def)) {
2932 found = true;
2933 break;
2934 }
2935 }
2936 if (!found) {
2937 assert(n != entry, "can't be entry");
2938 _data_entry.push(n);
2939 }
2940 }
2941 }
2942 }
2943
2944 // Find memory slices (head and tail)
2945 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2946 Node *n = lp()->fast_out(i);
2947 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2948 Node* n_tail = n->in(LoopNode::LoopBackControl);
2949 if (n_tail != n->in(LoopNode::EntryControl)) {
2950 if (!n_tail->is_Mem()) {
2951 assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2952 return false; // Bailout
2953 }
2954 _mem_slice_head.push(n);
2955 _mem_slice_tail.push(n_tail);
2956 }
2957 }
2958 }
2959
2960 // Create an RPO list of nodes in block
2961
2962 visited_clear();
2963 post_visited_clear();
2964
2965 // Push all non-control nodes with no inputs from within block, then control entry
2966 for (int j = 0; j < _data_entry.length(); j++) {
2967 Node* n = _data_entry.at(j);
2968 visited_set(n);
2969 _stk.push(n);
2970 }
2971 visited_set(entry);
2972 _stk.push(entry);
2973
2974 // Do a depth first walk over out edges
2975 int rpo_idx = bb_ct - 1;
2976 int size;
2977 int reduction_uses = 0;
2978 while ((size = _stk.length()) > 0) {
2979 Node* n = _stk.top(); // Leave node on stack
2980 if (!visited_test_set(n)) {
2981 // forward arc in graph
2982 } else if (!post_visited_test(n)) {
2983 // cross or back arc
2984 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2985 Node *use = n->fast_out(i);
2986 if (in_bb(use) && !visited_test(use) &&
2987 // Don't go around backedge
2988 (!use->is_Phi() || n == entry)) {
2989 if (use->is_reduction()) {
2990 // First see if we can map the reduction on the given system we are on, then
2991 // make a data entry operation for each reduction we see.
2992 BasicType bt = use->bottom_type()->basic_type();
2993 if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2994 reduction_uses++;
2995 }
2996 }
2997 _stk.push(use);
2998 }
2999 }
3000 if (_stk.length() == size) {
3001 // There were no additional uses, post visit node now
3002 _stk.pop(); // Remove node from stack
3003 assert(rpo_idx >= 0, "");
3004 _block.at_put_grow(rpo_idx, n);
3005 rpo_idx--;
3006 post_visited_set(n);
3007 assert(rpo_idx >= 0 || _stk.is_empty(), "");
3008 }
3009 } else {
3010 _stk.pop(); // Remove post-visited node from stack
3011 }
3012 }//while
3013
3014 int ii_current = -1;
3015 unsigned int load_idx = (unsigned int)-1;
3016 _ii_order.clear();
3017 // Create real map of block indices for nodes
3018 for (int j = 0; j < _block.length(); j++) {
3019 Node* n = _block.at(j);
3020 set_bb_idx(n, j);
3021 if (_do_vector_loop && n->is_Load()) {
3022 if (ii_current == -1) {
3023 ii_current = _clone_map.gen(n->_idx);
3024 _ii_order.push(ii_current);
3025 load_idx = _clone_map.idx(n->_idx);
3026 } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
3027 ii_current = _clone_map.gen(n->_idx);
3028 _ii_order.push(ii_current);
3029 }
3030 }
3031 }//for
3032
3033 // Ensure extra info is allocated.
3034 initialize_bb();
3035
3036 #ifndef PRODUCT
3037 if (_vector_loop_debug && _ii_order.length() > 0) {
3038 tty->print("SuperWord::construct_bb: List of generations: ");
3039 for (int jj = 0; jj < _ii_order.length(); ++jj) {
3040 tty->print(" %d:%d", jj, _ii_order.at(jj));
3041 }
3042 tty->print_cr(" ");
3043 }
3044 if (TraceSuperWord) {
3045 print_bb();
3046 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
3047 for (int m = 0; m < _data_entry.length(); m++) {
3048 tty->print("%3d ", m);
3049 _data_entry.at(m)->dump();
3050 }
3051 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
3052 for (int m = 0; m < _mem_slice_head.length(); m++) {
3053 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
3054 tty->print(" "); _mem_slice_tail.at(m)->dump();
3055 }
3056 }
3057 #endif
3058 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
3059 return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
3060 }
3061
3062 //------------------------------initialize_bb---------------------------
3063 // Initialize per node info
3064 void SuperWord::initialize_bb() {
3065 Node* last = _block.at(_block.length() - 1);
3066 grow_node_info(bb_idx(last));
3067 }
3068
3069 //------------------------------bb_insert_after---------------------------
3070 // Insert n into block after pos
3071 void SuperWord::bb_insert_after(Node* n, int pos) {
3072 int n_pos = pos + 1;
3073 // Make room
3074 for (int i = _block.length() - 1; i >= n_pos; i--) {
3075 _block.at_put_grow(i+1, _block.at(i));
3076 }
3077 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
3078 _node_info.at_put_grow(j+1, _node_info.at(j));
3079 }
3080 // Set value
3081 _block.at_put_grow(n_pos, n);
3082 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
3083 // Adjust map from node->_idx to _block index
3084 for (int i = n_pos; i < _block.length(); i++) {
3085 set_bb_idx(_block.at(i), i);
3086 }
3087 }
3088
3089 //------------------------------compute_max_depth---------------------------
3090 // Compute max depth for expressions from beginning of block
3091 // Use to prune search paths during test for independence.
3092 void SuperWord::compute_max_depth() {
3093 int ct = 0;
3094 bool again;
3095 do {
3096 again = false;
3097 for (int i = 0; i < _block.length(); i++) {
3098 Node* n = _block.at(i);
3099 if (!n->is_Phi()) {
3100 int d_orig = depth(n);
3101 int d_in = 0;
3102 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
3103 Node* pred = preds.current();
3104 if (in_bb(pred)) {
3105 d_in = MAX2(d_in, depth(pred));
3106 }
3107 }
3108 if (d_in + 1 != d_orig) {
3109 set_depth(n, d_in + 1);
3110 again = true;
3111 }
3112 }
3113 }
3114 ct++;
3115 } while (again);
3116
3117 if (TraceSuperWord && Verbose) {
3118 tty->print_cr("compute_max_depth iterated: %d times", ct);
3119 }
3120 }
3121
3122 //-------------------------compute_vector_element_type-----------------------
3123 // Compute necessary vector element type for expressions
3124 // This propagates backwards a narrower integer type when the
3125 // upper bits of the value are not needed.
3126 // Example: char a,b,c; a = b + c;
3127 // Normally the type of the add is integer, but for packed character
3128 // operations the type of the add needs to be char.
3129 void SuperWord::compute_vector_element_type() {
3130 if (TraceSuperWord && Verbose) {
3131 tty->print_cr("\ncompute_velt_type:");
3132 }
3133
3134 // Initial type
3135 for (int i = 0; i < _block.length(); i++) {
3136 Node* n = _block.at(i);
3137 set_velt_type(n, container_type(n));
3138 }
3139
3140 // Propagate integer narrowed type backwards through operations
3141 // that don't depend on higher order bits
3142 for (int i = _block.length() - 1; i >= 0; i--) {
3143 Node* n = _block.at(i);
3144 // Only integer types need be examined
3145 const Type* vtn = velt_type(n);
3146 if (vtn->basic_type() == T_INT) {
3147 uint start, end;
3148 VectorNode::vector_operands(n, &start, &end);
3149
3150 for (uint j = start; j < end; j++) {
3151 Node* in = n->in(j);
3152 // Don't propagate through a memory
3153 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
3154 data_size(n) < data_size(in)) {
3155 bool same_type = true;
3156 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
3157 Node *use = in->fast_out(k);
3158 if (!in_bb(use) || !same_velt_type(use, n)) {
3159 same_type = false;
3160 break;
3161 }
3162 }
3163 if (same_type) {
3164 // For right shifts of small integer types (bool, byte, char, short)
3165 // we need precise information about sign-ness. Only Load nodes have
3166 // this information because Store nodes are the same for signed and
3167 // unsigned values. And any arithmetic operation after a load may
3168 // expand a value to signed Int so such right shifts can't be used
3169 // because vector elements do not have upper bits of Int.
3170 const Type* vt = vtn;
3171 if (VectorNode::is_shift(in)) {
3172 Node* load = in->in(1);
3173 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
3174 vt = velt_type(load);
3175 } else if (in->Opcode() != Op_LShiftI) {
3176 // Widen type to Int to avoid creation of right shift vector
3177 // (align + data_size(s1) check in stmts_can_pack() will fail).
3178 // Note, left shifts work regardless type.
3179 vt = TypeInt::INT;
3180 }
3181 }
3182 set_velt_type(in, vt);
3183 }
3184 }
3185 }
3186 }
3187 }
3188 #ifndef PRODUCT
3189 if (TraceSuperWord && Verbose) {
3190 for (int i = 0; i < _block.length(); i++) {
3191 Node* n = _block.at(i);
3192 velt_type(n)->dump();
3193 tty->print("\t");
3194 n->dump();
3195 }
3196 }
3197 #endif
3198 }
3199
3200 //------------------------------memory_alignment---------------------------
3201 // Alignment within a vector memory reference
3202 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
3203 #ifndef PRODUCT
3204 if(TraceSuperWord && Verbose) {
3205 tty->print("SuperWord::memory_alignment within a vector memory reference for %d: ", s->_idx); s->dump();
3206 }
3207 #endif
3208 NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
3209 SWPointer p(s, this, NULL, false);
3210 if (!p.valid()) {
3211 NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
3212 return bottom_align;
3213 }
3214 int vw = get_vw_bytes_special(s);
3215 if (vw < 2) {
3216 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
3217 return bottom_align; // No vectors for this type
3218 }
3219 int offset = p.offset_in_bytes();
3220 offset += iv_adjust*p.memory_size();
3221 int off_rem = offset % vw;
3222 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
3223 if (TraceSuperWord && Verbose) {
3224 tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
3225 }
3226 return off_mod;
3227 }
3228
3229 //---------------------------container_type---------------------------
3230 // Smallest type containing range of values
3231 const Type* SuperWord::container_type(Node* n) {
3232 if (n->is_Mem()) {
3233 BasicType bt = n->as_Mem()->memory_type();
3234 if (n->is_Store() && (bt == T_CHAR)) {
3235 // Use T_SHORT type instead of T_CHAR for stored values because any
3236 // preceding arithmetic operation extends values to signed Int.
3237 bt = T_SHORT;
3238 }
3239 if (n->Opcode() == Op_LoadUB) {
3240 // Adjust type for unsigned byte loads, it is important for right shifts.
3241 // T_BOOLEAN is used because there is no basic type representing type
3242 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
3243 // size (one byte) and sign is important.
3244 bt = T_BOOLEAN;
3245 }
3246 return Type::get_const_basic_type(bt);
3247 }
3248 const Type* t = _igvn.type(n);
3249 if (t->basic_type() == T_INT) {
3250 // A narrow type of arithmetic operations will be determined by
3251 // propagating the type of memory operations.
3252 return TypeInt::INT;
3253 }
3254 return t;
3255 }
3256
3257 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
3258 const Type* vt1 = velt_type(n1);
3259 const Type* vt2 = velt_type(n2);
3260 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
3261 // Compare vectors element sizes for integer types.
3262 return data_size(n1) == data_size(n2);
3263 }
3264 return vt1 == vt2;
3265 }
3266
3267 //------------------------------in_packset---------------------------
3268 // Are s1 and s2 in a pack pair and ordered as s1,s2?
3269 bool SuperWord::in_packset(Node* s1, Node* s2) {
3270 for (int i = 0; i < _packset.length(); i++) {
3271 Node_List* p = _packset.at(i);
3272 assert(p->size() == 2, "must be");
3273 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
3274 return true;
3275 }
3276 }
3277 return false;
3278 }
3279
3280 //------------------------------in_pack---------------------------
3281 // Is s in pack p?
3282 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
3283 for (uint i = 0; i < p->size(); i++) {
3284 if (p->at(i) == s) {
3285 return p;
3286 }
3287 }
3288 return NULL;
3289 }
3290
3291 //------------------------------remove_pack_at---------------------------
3292 // Remove the pack at position pos in the packset
3293 void SuperWord::remove_pack_at(int pos) {
3294 Node_List* p = _packset.at(pos);
3295 for (uint i = 0; i < p->size(); i++) {
3296 Node* s = p->at(i);
3297 set_my_pack(s, NULL);
3298 }
3299 _packset.remove_at(pos);
3300 }
3301
3302 void SuperWord::packset_sort(int n) {
3303 // simple bubble sort so that we capitalize with O(n) when its already sorted
3304 while (n != 0) {
3305 bool swapped = false;
3306 for (int i = 1; i < n; i++) {
3307 Node_List* q_low = _packset.at(i-1);
3308 Node_List* q_i = _packset.at(i);
3309
3310 // only swap when we find something to swap
3311 if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
3312 Node_List* t = q_i;
3313 *(_packset.adr_at(i)) = q_low;
3314 *(_packset.adr_at(i-1)) = q_i;
3315 swapped = true;
3316 }
3317 }
3318 if (swapped == false) break;
3319 n--;
3320 }
3321 }
3322
3323 //------------------------------executed_first---------------------------
3324 // Return the node executed first in pack p. Uses the RPO block list
3325 // to determine order.
3326 Node* SuperWord::executed_first(Node_List* p) {
3327 Node* n = p->at(0);
3328 int n_rpo = bb_idx(n);
3329 for (uint i = 1; i < p->size(); i++) {
3330 Node* s = p->at(i);
3331 int s_rpo = bb_idx(s);
3332 if (s_rpo < n_rpo) {
3333 n = s;
3334 n_rpo = s_rpo;
3335 }
3336 }
3337 return n;
3338 }
3339
3340 //------------------------------executed_last---------------------------
3341 // Return the node executed last in pack p.
3342 Node* SuperWord::executed_last(Node_List* p) {
3343 Node* n = p->at(0);
3344 int n_rpo = bb_idx(n);
3345 for (uint i = 1; i < p->size(); i++) {
3346 Node* s = p->at(i);
3347 int s_rpo = bb_idx(s);
3348 if (s_rpo > n_rpo) {
3349 n = s;
3350 n_rpo = s_rpo;
3351 }
3352 }
3353 return n;
3354 }
3355
3356 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
3357 LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
3358 for (uint i = 0; i < p->size(); i++) {
3359 Node* n = p->at(i);
3360 assert(n->is_Load(), "only meaningful for loads");
3361 if (!n->depends_only_on_test()) {
3362 if (n->as_Load()->has_unknown_control_dependency() &&
3363 dep != LoadNode::Pinned) {
3364 // Upgrade to unknown control...
3365 dep = LoadNode::UnknownControl;
3366 } else {
3367 // Otherwise, we must pin it.
3368 dep = LoadNode::Pinned;
3369 }
3370 }
3371 }
3372 return dep;
3373 }
3374
3375
3376 //----------------------------align_initial_loop_index---------------------------
3377 // Adjust pre-loop limit so that in main loop, a load/store reference
3378 // to align_to_ref will be a position zero in the vector.
3379 // (iv + k) mod vector_align == 0
3380 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
3381 CountedLoopNode *main_head = lp()->as_CountedLoop();
3382 assert(main_head->is_main_loop(), "");
3383 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
3384 assert(pre_end != NULL, "we must have a correct pre-loop");
3385 Node *pre_opaq1 = pre_end->limit();
3386 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
3387 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
3388 Node *lim0 = pre_opaq->in(1);
3389
3390 // Where we put new limit calculations
3391 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
3392
3393 // Ensure the original loop limit is available from the
3394 // pre-loop Opaque1 node.
3395 Node *orig_limit = pre_opaq->original_loop_limit();
3396 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
3397
3398 SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
3399 assert(align_to_ref_p.valid(), "sanity");
3400
3401 // Given:
3402 // lim0 == original pre loop limit
3403 // V == v_align (power of 2)
3404 // invar == extra invariant piece of the address expression
3405 // e == offset [ +/- invar ]
3406 //
3407 // When reassociating expressions involving '%' the basic rules are:
3408 // (a - b) % k == 0 => a % k == b % k
3409 // and:
3410 // (a + b) % k == 0 => a % k == (k - b) % k
3411 //
3412 // For stride > 0 && scale > 0,
3413 // Derive the new pre-loop limit "lim" such that the two constraints:
3414 // (1) lim = lim0 + N (where N is some positive integer < V)
3415 // (2) (e + lim) % V == 0
3416 // are true.
3417 //
3418 // Substituting (1) into (2),
3419 // (e + lim0 + N) % V == 0
3420 // solve for N:
3421 // N = (V - (e + lim0)) % V
3422 // substitute back into (1), so that new limit
3423 // lim = lim0 + (V - (e + lim0)) % V
3424 //
3425 // For stride > 0 && scale < 0
3426 // Constraints:
3427 // lim = lim0 + N
3428 // (e - lim) % V == 0
3429 // Solving for lim:
3430 // (e - lim0 - N) % V == 0
3431 // N = (e - lim0) % V
3432 // lim = lim0 + (e - lim0) % V
3433 //
3434 // For stride < 0 && scale > 0
3435 // Constraints:
3436 // lim = lim0 - N
3437 // (e + lim) % V == 0
3438 // Solving for lim:
3439 // (e + lim0 - N) % V == 0
3440 // N = (e + lim0) % V
3441 // lim = lim0 - (e + lim0) % V
3442 //
3443 // For stride < 0 && scale < 0
3444 // Constraints:
3445 // lim = lim0 - N
3446 // (e - lim) % V == 0
3447 // Solving for lim:
3448 // (e - lim0 + N) % V == 0
3449 // N = (V - (e - lim0)) % V
3450 // lim = lim0 - (V - (e - lim0)) % V
3451
3452 int vw = vector_width_in_bytes(align_to_ref);
3453 int stride = iv_stride();
3454 int scale = align_to_ref_p.scale_in_bytes();
3455 int elt_size = align_to_ref_p.memory_size();
3456 int v_align = vw / elt_size;
3457 assert(v_align > 1, "sanity");
3458 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
3459 Node *offsn = _igvn.intcon(offset);
3460
3461 Node *e = offsn;
3462 if (align_to_ref_p.invar() != NULL) {
3463 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3464 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3465 Node* invar = align_to_ref_p.invar();
3466 if (_igvn.type(invar)->isa_long()) {
3467 // Computations are done % (vector width/element size) so it's
3468 // safe to simply convert invar to an int and loose the upper 32
3469 // bit half.
3470 invar = new ConvL2INode(invar);
3471 _igvn.register_new_node_with_optimizer(invar);
3472 }
3473 Node* aref = new URShiftINode(invar, log2_elt);
3474 _igvn.register_new_node_with_optimizer(aref);
3475 _phase->set_ctrl(aref, pre_ctrl);
3476 if (align_to_ref_p.negate_invar()) {
3477 e = new SubINode(e, aref);
3478 } else {
3479 e = new AddINode(e, aref);
3480 }
3481 _igvn.register_new_node_with_optimizer(e);
3482 _phase->set_ctrl(e, pre_ctrl);
3483 }
3484 if (vw > ObjectAlignmentInBytes || align_to_ref_p.base()->is_top()) {
3485 // incorporate base e +/- base && Mask >>> log2(elt)
3486 Node* xbase = new CastP2XNode(NULL, align_to_ref_p.adr());
3487 _igvn.register_new_node_with_optimizer(xbase);
3488 #ifdef _LP64
3489 xbase = new ConvL2INode(xbase);
3490 _igvn.register_new_node_with_optimizer(xbase);
3491 #endif
3492 Node* mask = _igvn.intcon(vw-1);
3493 Node* masked_xbase = new AndINode(xbase, mask);
3494 _igvn.register_new_node_with_optimizer(masked_xbase);
3495 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3496 Node* bref = new URShiftINode(masked_xbase, log2_elt);
3497 _igvn.register_new_node_with_optimizer(bref);
3498 _phase->set_ctrl(bref, pre_ctrl);
3499 e = new AddINode(e, bref);
3500 _igvn.register_new_node_with_optimizer(e);
3501 _phase->set_ctrl(e, pre_ctrl);
3502 }
3503
3504 // compute e +/- lim0
3505 if (scale < 0) {
3506 e = new SubINode(e, lim0);
3507 } else {
3508 e = new AddINode(e, lim0);
3509 }
3510 _igvn.register_new_node_with_optimizer(e);
3511 _phase->set_ctrl(e, pre_ctrl);
3512
3513 if (stride * scale > 0) {
3514 // compute V - (e +/- lim0)
3515 Node* va = _igvn.intcon(v_align);
3516 e = new SubINode(va, e);
3517 _igvn.register_new_node_with_optimizer(e);
3518 _phase->set_ctrl(e, pre_ctrl);
3519 }
3520 // compute N = (exp) % V
3521 Node* va_msk = _igvn.intcon(v_align - 1);
3522 Node* N = new AndINode(e, va_msk);
3523 _igvn.register_new_node_with_optimizer(N);
3524 _phase->set_ctrl(N, pre_ctrl);
3525
3526 // substitute back into (1), so that new limit
3527 // lim = lim0 + N
3528 Node* lim;
3529 if (stride < 0) {
3530 lim = new SubINode(lim0, N);
3531 } else {
3532 lim = new AddINode(lim0, N);
3533 }
3534 _igvn.register_new_node_with_optimizer(lim);
3535 _phase->set_ctrl(lim, pre_ctrl);
3536 Node* constrained =
3537 (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3538 : (Node*) new MaxINode(lim, orig_limit);
3539 _igvn.register_new_node_with_optimizer(constrained);
3540 _phase->set_ctrl(constrained, pre_ctrl);
3541 _igvn.replace_input_of(pre_opaq, 1, constrained);
3542 }
3543
3544 //----------------------------get_pre_loop_end---------------------------
3545 // Find pre loop end from main loop. Returns null if none.
3546 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3547 // The loop cannot be optimized if the graph shape at
3548 // the loop entry is inappropriate.
3549 if (!PhaseIdealLoop::is_canonical_loop_entry(cl)) {
3550 return NULL;
3551 }
3552
3553 Node* p_f = cl->skip_predicates()->in(0)->in(0);
3554 if (!p_f->is_IfFalse()) return NULL;
3555 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3556 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3557 CountedLoopNode* loop_node = pre_end->loopnode();
3558 if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3559 return pre_end;
3560 }
3561
3562 //------------------------------init---------------------------
3563 void SuperWord::init() {
3564 _dg.init();
3565 _packset.clear();
3566 _disjoint_ptrs.clear();
3567 _block.clear();
3568 _post_block.clear();
3569 _data_entry.clear();
3570 _mem_slice_head.clear();
3571 _mem_slice_tail.clear();
3572 _iteration_first.clear();
3573 _iteration_last.clear();
3574 _node_info.clear();
3575 _align_to_ref = NULL;
3576 _lpt = NULL;
3577 _lp = NULL;
3578 _bb = NULL;
3579 _iv = NULL;
3580 _race_possible = 0;
3581 _early_return = false;
3582 _num_work_vecs = 0;
3583 _num_reductions = 0;
3584 }
3585
3586 //------------------------------restart---------------------------
3587 void SuperWord::restart() {
3588 _dg.init();
3589 _packset.clear();
3590 _disjoint_ptrs.clear();
3591 _block.clear();
3592 _post_block.clear();
3593 _data_entry.clear();
3594 _mem_slice_head.clear();
3595 _mem_slice_tail.clear();
3596 _node_info.clear();
3597 }
3598
3599 //------------------------------print_packset---------------------------
3600 void SuperWord::print_packset() {
3601 #ifndef PRODUCT
3602 tty->print_cr("packset");
3603 for (int i = 0; i < _packset.length(); i++) {
3604 tty->print_cr("Pack: %d", i);
3605 Node_List* p = _packset.at(i);
3606 print_pack(p);
3607 }
3608 #endif
3609 }
3610
3611 //------------------------------print_pack---------------------------
3612 void SuperWord::print_pack(Node_List* p) {
3613 for (uint i = 0; i < p->size(); i++) {
3614 print_stmt(p->at(i));
3615 }
3616 }
3617
3618 //------------------------------print_bb---------------------------
3619 void SuperWord::print_bb() {
3620 #ifndef PRODUCT
3621 tty->print_cr("\nBlock");
3622 for (int i = 0; i < _block.length(); i++) {
3623 Node* n = _block.at(i);
3624 tty->print("%d ", i);
3625 if (n) {
3626 n->dump();
3627 }
3628 }
3629 #endif
3630 }
3631
3632 //------------------------------print_stmt---------------------------
3633 void SuperWord::print_stmt(Node* s) {
3634 #ifndef PRODUCT
3635 tty->print(" align: %d \t", alignment(s));
3636 s->dump();
3637 #endif
3638 }
3639
3640 //------------------------------blank---------------------------
3641 char* SuperWord::blank(uint depth) {
3642 static char blanks[101];
3643 assert(depth < 101, "too deep");
3644 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3645 blanks[depth] = '\0';
3646 return blanks;
3647 }
3648
3649
3650 //==============================SWPointer===========================
3651 #ifndef PRODUCT
3652 int SWPointer::Tracer::_depth = 0;
3653 #endif
3654 //----------------------------SWPointer------------------------
3655 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3656 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
3657 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3658 _nstack(nstack), _analyze_only(analyze_only),
3659 _stack_idx(0)
3660 #ifndef PRODUCT
3661 , _tracer(slp)
3662 #endif
3663 {
3664 NOT_PRODUCT(_tracer.ctor_1(mem);)
3665
3666 Node* adr = mem->in(MemNode::Address);
3667 if (!adr->is_AddP()) {
3668 assert(!valid(), "too complex");
3669 return;
3670 }
3671 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3672 Node* base = adr->in(AddPNode::Base);
3673 // The base address should be loop invariant
3674 if (!invariant(base)) {
3675 assert(!valid(), "base address is loop variant");
3676 return;
3677 }
3678 // unsafe references require misaligned vector access support
3679 if (base->is_top() && !Matcher::misaligned_vectors_ok()) {
3680 assert(!valid(), "unsafe access");
3681 return;
3682 }
3683
3684 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3685 NOT_PRODUCT(_tracer.ctor_2(adr);)
3686
3687 int i;
3688 for (i = 0; i < 3; i++) {
3689 NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3690
3691 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3692 assert(!valid(), "too complex");
3693 return;
3694 }
3695 adr = adr->in(AddPNode::Address);
3696 NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3697
3698 if (base == adr || !adr->is_AddP()) {
3699 NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3700 break; // stop looking at addp's
3701 }
3702 }
3703 if (!invariant(adr)) {
3704 assert(!valid(), "adr is loop variant");
3705 return;
3706 }
3707
3708 if (!base->is_top() && adr != base) {
3709 assert(!valid(), "adr and base differ");
3710 return;
3711 }
3712
3713 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3714 NOT_PRODUCT(_tracer.ctor_6(mem);)
3715
3716 _base = base;
3717 _adr = adr;
3718 assert(valid(), "Usable");
3719 }
3720
3721 // Following is used to create a temporary object during
3722 // the pattern match of an address expression.
3723 SWPointer::SWPointer(SWPointer* p) :
3724 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
3725 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3726 _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3727 _stack_idx(p->_stack_idx)
3728 #ifndef PRODUCT
3729 , _tracer(p->_slp)
3730 #endif
3731 {}
3732
3733
3734 bool SWPointer::invariant(Node* n) {
3735 NOT_PRODUCT(Tracer::Depth dd;)
3736 Node *n_c = phase()->get_ctrl(n);
3737 NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3738 return !lpt()->is_member(phase()->get_loop(n_c));
3739 }
3740 //------------------------scaled_iv_plus_offset--------------------
3741 // Match: k*iv + offset
3742 // where: k is a constant that maybe zero, and
3743 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3744 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3745 NOT_PRODUCT(Tracer::Depth ddd;)
3746 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3747
3748 if (scaled_iv(n)) {
3749 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3750 return true;
3751 }
3752
3753 if (offset_plus_k(n)) {
3754 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3755 return true;
3756 }
3757
3758 int opc = n->Opcode();
3759 if (opc == Op_AddI) {
3760 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3761 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3762 return true;
3763 }
3764 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3765 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3766 return true;
3767 }
3768 } else if (opc == Op_SubI) {
3769 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3770 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3771 return true;
3772 }
3773 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3774 _scale *= -1;
3775 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3776 return true;
3777 }
3778 }
3779
3780 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3781 return false;
3782 }
3783
3784 //----------------------------scaled_iv------------------------
3785 // Match: k*iv where k is a constant that's not zero
3786 bool SWPointer::scaled_iv(Node* n) {
3787 NOT_PRODUCT(Tracer::Depth ddd;)
3788 NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3789
3790 if (_scale != 0) { // already found a scale
3791 NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3792 return false;
3793 }
3794
3795 if (n == iv()) {
3796 _scale = 1;
3797 NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3798 return true;
3799 }
3800 if (_analyze_only && (invariant(n) == false)) {
3801 _nstack->push(n, _stack_idx++);
3802 }
3803
3804 int opc = n->Opcode();
3805 if (opc == Op_MulI) {
3806 if (n->in(1) == iv() && n->in(2)->is_Con()) {
3807 _scale = n->in(2)->get_int();
3808 NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3809 return true;
3810 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3811 _scale = n->in(1)->get_int();
3812 NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3813 return true;
3814 }
3815 } else if (opc == Op_LShiftI) {
3816 if (n->in(1) == iv() && n->in(2)->is_Con()) {
3817 _scale = 1 << n->in(2)->get_int();
3818 NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3819 return true;
3820 }
3821 } else if (opc == Op_ConvI2L) {
3822 if (n->in(1)->Opcode() == Op_CastII &&
3823 n->in(1)->as_CastII()->has_range_check()) {
3824 // Skip range check dependent CastII nodes
3825 n = n->in(1);
3826 }
3827 if (scaled_iv_plus_offset(n->in(1))) {
3828 NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3829 return true;
3830 }
3831 } else if (opc == Op_LShiftL) {
3832 if (!has_iv() && _invar == NULL) {
3833 // Need to preserve the current _offset value, so
3834 // create a temporary object for this expression subtree.
3835 // Hacky, so should re-engineer the address pattern match.
3836 NOT_PRODUCT(Tracer::Depth dddd;)
3837 SWPointer tmp(this);
3838 NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3839
3840 if (tmp.scaled_iv_plus_offset(n->in(1))) {
3841 if (tmp._invar == NULL || _slp->do_vector_loop()) {
3842 int mult = 1 << n->in(2)->get_int();
3843 _scale = tmp._scale * mult;
3844 _offset += tmp._offset * mult;
3845 NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3846 return true;
3847 }
3848 }
3849 }
3850 }
3851 NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3852 return false;
3853 }
3854
3855 //----------------------------offset_plus_k------------------------
3856 // Match: offset is (k [+/- invariant])
3857 // where k maybe zero and invariant is optional, but not both.
3858 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3859 NOT_PRODUCT(Tracer::Depth ddd;)
3860 NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3861
3862 int opc = n->Opcode();
3863 if (opc == Op_ConI) {
3864 _offset += negate ? -(n->get_int()) : n->get_int();
3865 NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3866 return true;
3867 } else if (opc == Op_ConL) {
3868 // Okay if value fits into an int
3869 const TypeLong* t = n->find_long_type();
3870 if (t->higher_equal(TypeLong::INT)) {
3871 jlong loff = n->get_long();
3872 jint off = (jint)loff;
3873 _offset += negate ? -off : loff;
3874 NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3875 return true;
3876 }
3877 NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3878 return false;
3879 }
3880 if (_invar != NULL) { // already has an invariant
3881 NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3882 return false;
3883 }
3884
3885 if (_analyze_only && (invariant(n) == false)) {
3886 _nstack->push(n, _stack_idx++);
3887 }
3888 if (opc == Op_AddI) {
3889 if (n->in(2)->is_Con() && invariant(n->in(1))) {
3890 _negate_invar = negate;
3891 _invar = n->in(1);
3892 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3893 NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3894 return true;
3895 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3896 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3897 _negate_invar = negate;
3898 _invar = n->in(2);
3899 NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3900 return true;
3901 }
3902 }
3903 if (opc == Op_SubI) {
3904 if (n->in(2)->is_Con() && invariant(n->in(1))) {
3905 _negate_invar = negate;
3906 _invar = n->in(1);
3907 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3908 NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3909 return true;
3910 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3911 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3912 _negate_invar = !negate;
3913 _invar = n->in(2);
3914 NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3915 return true;
3916 }
3917 }
3918 if (invariant(n)) {
3919 if (opc == Op_ConvI2L) {
3920 n = n->in(1);
3921 if (n->Opcode() == Op_CastII &&
3922 n->as_CastII()->has_range_check()) {
3923 // Skip range check dependent CastII nodes
3924 assert(invariant(n), "sanity");
3925 n = n->in(1);
3926 }
3927 }
3928 _negate_invar = negate;
3929 _invar = n;
3930 NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3931 return true;
3932 }
3933
3934 NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3935 return false;
3936 }
3937
3938 //----------------------------print------------------------
3939 void SWPointer::print() {
3940 #ifndef PRODUCT
3941 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
3942 _base != NULL ? _base->_idx : 0,
3943 _adr != NULL ? _adr->_idx : 0,
3944 _scale, _offset,
3945 _negate_invar?'-':'+',
3946 _invar != NULL ? _invar->_idx : 0);
3947 #endif
3948 }
3949
3950 //----------------------------tracing------------------------
3951 #ifndef PRODUCT
3952 void SWPointer::Tracer::print_depth() {
3953 for (int ii = 0; ii<_depth; ++ii) tty->print(" ");
3954 }
3955
3956 void SWPointer::Tracer::ctor_1 (Node* mem) {
3957 if(_slp->is_trace_alignment()) {
3958 print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3959 }
3960 }
3961
3962 void SWPointer::Tracer::ctor_2(Node* adr) {
3963 if(_slp->is_trace_alignment()) {
3964 //store_depth();
3965 inc_depth();
3966 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3967 inc_depth();
3968 print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3969 }
3970 }
3971
3972 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3973 if(_slp->is_trace_alignment()) {
3974 inc_depth();
3975 Node* offset = adr->in(AddPNode::Offset);
3976 print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3977 }
3978 }
3979
3980 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3981 if(_slp->is_trace_alignment()) {
3982 inc_depth();
3983 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3984 }
3985 }
3986
3987 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3988 if(_slp->is_trace_alignment()) {
3989 inc_depth();
3990 if (base == adr) {
3991 print_depth(); tty->print_cr(" \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3992 } else if (!adr->is_AddP()) {
3993 print_depth(); tty->print_cr(" \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
3994 }
3995 }
3996 }
3997
3998 void SWPointer::Tracer::ctor_6(Node* mem) {
3999 if(_slp->is_trace_alignment()) {
4000 //restore_depth();
4001 print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
4002 }
4003 }
4004
4005 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
4006 if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
4007 int is_member = _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
4008 int in_bb = _slp->in_bb(n);
4009 print_depth(); tty->print(" \\ "); tty->print_cr(" %d SWPointer::invariant conditions differ: n_c %d", n->_idx, n_c->_idx);
4010 print_depth(); tty->print(" \\ "); tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
4011 print_depth(); tty->print(" \\ "); n->dump();
4012 print_depth(); tty->print(" \\ "); n_c->dump();
4013 }
4014 }
4015
4016 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
4017 if(_slp->is_trace_alignment()) {
4018 print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
4019 n->dump();
4020 }
4021 }
4022
4023 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
4024 if(_slp->is_trace_alignment()) {
4025 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
4026 }
4027 }
4028
4029 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
4030 if(_slp->is_trace_alignment()) {
4031 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
4032 }
4033 }
4034
4035 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
4036 if(_slp->is_trace_alignment()) {
4037 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
4038 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
4039 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
4040 }
4041 }
4042
4043 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
4044 if(_slp->is_trace_alignment()) {
4045 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
4046 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
4047 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
4048 }
4049 }
4050
4051 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
4052 if(_slp->is_trace_alignment()) {
4053 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
4054 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
4055 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
4056 }
4057 }
4058
4059 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
4060 if(_slp->is_trace_alignment()) {
4061 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
4062 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
4063 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
4064 }
4065 }
4066
4067 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
4068 if(_slp->is_trace_alignment()) {
4069 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
4070 }
4071 }
4072
4073 void SWPointer::Tracer::scaled_iv_1(Node* n) {
4074 if(_slp->is_trace_alignment()) {
4075 print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
4076 }
4077 }
4078
4079 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
4080 if(_slp->is_trace_alignment()) {
4081 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
4082 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
4083 }
4084 }
4085
4086 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
4087 if(_slp->is_trace_alignment()) {
4088 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
4089 }
4090 }
4091
4092 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
4093 if(_slp->is_trace_alignment()) {
4094 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
4095 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
4096 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4097 }
4098 }
4099
4100 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
4101 if(_slp->is_trace_alignment()) {
4102 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
4103 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
4104 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4105 }
4106 }
4107
4108 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
4109 if(_slp->is_trace_alignment()) {
4110 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
4111 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
4112 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4113 }
4114 }
4115
4116 void SWPointer::Tracer::scaled_iv_7(Node* n) {
4117 if(_slp->is_trace_alignment()) {
4118 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
4119 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
4120 inc_depth(); inc_depth();
4121 print_depth(); n->in(1)->dump();
4122 dec_depth(); dec_depth();
4123 }
4124 }
4125
4126 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
4127 if(_slp->is_trace_alignment()) {
4128 print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
4129 }
4130 }
4131
4132 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
4133 if(_slp->is_trace_alignment()) {
4134 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
4135 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",
4136 n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
4137 inc_depth(); inc_depth();
4138 print_depth(); n->in(1)->dump();
4139 print_depth(); n->in(2)->dump();
4140 dec_depth(); dec_depth();
4141 }
4142 }
4143
4144 void SWPointer::Tracer::scaled_iv_10(Node* n) {
4145 if(_slp->is_trace_alignment()) {
4146 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
4147 }
4148 }
4149
4150 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
4151 if(_slp->is_trace_alignment()) {
4152 print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
4153 }
4154 }
4155
4156 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
4157 if(_slp->is_trace_alignment()) {
4158 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
4159 }
4160 }
4161
4162 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
4163 if(_slp->is_trace_alignment()) {
4164 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
4165 }
4166 }
4167
4168 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
4169 if(_slp->is_trace_alignment()) {
4170 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4171 print_depth(); tty->print_cr(" \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
4172 }
4173 }
4174
4175 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
4176 if(_slp->is_trace_alignment()) {
4177 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
4178 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
4179 }
4180 }
4181
4182 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4183 if(_slp->is_trace_alignment()) {
4184 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4185 n->_idx, _negate_invar, _invar->_idx, _offset);
4186 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4187 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4188 }
4189 }
4190
4191 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4192 if(_slp->is_trace_alignment()) {
4193 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4194 n->_idx, _negate_invar, _invar->_idx, _offset);
4195 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4196 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4197 }
4198 }
4199
4200 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4201 if(_slp->is_trace_alignment()) {
4202 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4203 n->_idx, _negate_invar, _invar->_idx, _offset);
4204 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4205 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4206 }
4207 }
4208
4209 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4210 if(_slp->is_trace_alignment()) {
4211 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);
4212 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4213 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4214 }
4215 }
4216
4217 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4218 if(_slp->is_trace_alignment()) {
4219 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);
4220 print_depth(); tty->print_cr(" \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
4221 }
4222 }
4223
4224 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
4225 if(_slp->is_trace_alignment()) {
4226 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4227 }
4228 }
4229
4230 #endif
4231 // ========================= OrderedPair =====================
4232
4233 const OrderedPair OrderedPair::initial;
4234
4235 // ========================= SWNodeInfo =====================
4236
4237 const SWNodeInfo SWNodeInfo::initial;
4238
4239
4240 // ============================ DepGraph ===========================
4241
4242 //------------------------------make_node---------------------------
4243 // Make a new dependence graph node for an ideal node.
4244 DepMem* DepGraph::make_node(Node* node) {
4245 DepMem* m = new (_arena) DepMem(node);
4246 if (node != NULL) {
4247 assert(_map.at_grow(node->_idx) == NULL, "one init only");
4248 _map.at_put_grow(node->_idx, m);
4249 }
4250 return m;
4251 }
4252
4253 //------------------------------make_edge---------------------------
4254 // Make a new dependence graph edge from dpred -> dsucc
4255 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
4256 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
4257 dpred->set_out_head(e);
4258 dsucc->set_in_head(e);
4259 return e;
4260 }
4261
4262 // ========================== DepMem ========================
4263
4264 //------------------------------in_cnt---------------------------
4265 int DepMem::in_cnt() {
4266 int ct = 0;
4267 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
4268 return ct;
4269 }
4270
4271 //------------------------------out_cnt---------------------------
4272 int DepMem::out_cnt() {
4273 int ct = 0;
4274 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
4275 return ct;
4276 }
4277
4278 //------------------------------print-----------------------------
4279 void DepMem::print() {
4280 #ifndef PRODUCT
4281 tty->print(" DepNode %d (", _node->_idx);
4282 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
4283 Node* pred = p->pred()->node();
4284 tty->print(" %d", pred != NULL ? pred->_idx : 0);
4285 }
4286 tty->print(") [");
4287 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
4288 Node* succ = s->succ()->node();
4289 tty->print(" %d", succ != NULL ? succ->_idx : 0);
4290 }
4291 tty->print_cr(" ]");
4292 #endif
4293 }
4294
4295 // =========================== DepEdge =========================
4296
4297 //------------------------------DepPreds---------------------------
4298 void DepEdge::print() {
4299 #ifndef PRODUCT
4300 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
4301 #endif
4302 }
4303
4304 // =========================== DepPreds =========================
4305 // Iterator over predecessor edges in the dependence graph.
4306
4307 //------------------------------DepPreds---------------------------
4308 DepPreds::DepPreds(Node* n, DepGraph& dg) {
4309 _n = n;
4310 _done = false;
4311 if (_n->is_Store() || _n->is_Load()) {
4312 _next_idx = MemNode::Address;
4313 _end_idx = n->req();
4314 _dep_next = dg.dep(_n)->in_head();
4315 } else if (_n->is_Mem()) {
4316 _next_idx = 0;
4317 _end_idx = 0;
4318 _dep_next = dg.dep(_n)->in_head();
4319 } else {
4320 _next_idx = 1;
4321 _end_idx = _n->req();
4322 _dep_next = NULL;
4323 }
4324 next();
4325 }
4326
4327 //------------------------------next---------------------------
4328 void DepPreds::next() {
4329 if (_dep_next != NULL) {
4330 _current = _dep_next->pred()->node();
4331 _dep_next = _dep_next->next_in();
4332 } else if (_next_idx < _end_idx) {
4333 _current = _n->in(_next_idx++);
4334 } else {
4335 _done = true;
4336 }
4337 }
4338
4339 // =========================== DepSuccs =========================
4340 // Iterator over successor edges in the dependence graph.
4341
4342 //------------------------------DepSuccs---------------------------
4343 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
4344 _n = n;
4345 _done = false;
4346 if (_n->is_Load()) {
4347 _next_idx = 0;
4348 _end_idx = _n->outcnt();
4349 _dep_next = dg.dep(_n)->out_head();
4350 } else if (_n->is_Mem() || (_n->is_Phi() && _n->bottom_type() == Type::MEMORY)) {
4351 _next_idx = 0;
4352 _end_idx = 0;
4353 _dep_next = dg.dep(_n)->out_head();
4354 } else {
4355 _next_idx = 0;
4356 _end_idx = _n->outcnt();
4357 _dep_next = NULL;
4358 }
4359 next();
4360 }
4361
4362 //-------------------------------next---------------------------
4363 void DepSuccs::next() {
4364 if (_dep_next != NULL) {
4365 _current = _dep_next->succ()->node();
4366 _dep_next = _dep_next->next_out();
4367 } else if (_next_idx < _end_idx) {
4368 _current = _n->raw_out(_next_idx++);
4369 } else {
4370 _done = true;
4371 }
4372 }
4373
4374 //
4375 // --------------------------------- vectorization/simd -----------------------------------
4376 //
4377 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
4378 return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
4379 }
4380 bool SuperWord::same_generation(Node* a, Node* b) const {
4381 return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
4382 }
4383
4384 Node* SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
4385 assert(in_bb(ld), "must be in block");
4386 if (_clone_map.gen(ld->_idx) == _ii_first) {
4387 #ifndef PRODUCT
4388 if (_vector_loop_debug) {
4389 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
4390 _clone_map.gen(ld->_idx));
4391 }
4392 #endif
4393 return NULL; //we think that any ld in the first gen being vectorizable
4394 }
4395
4396 Node* mem = ld->in(MemNode::Memory);
4397 if (mem->outcnt() <= 1) {
4398 // we don't want to remove the only edge from mem node to load
4399 #ifndef PRODUCT
4400 if (_vector_loop_debug) {
4401 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",
4402 mem->_idx, ld->_idx);
4403 ld->dump();
4404 mem->dump();
4405 }
4406 #endif
4407 return NULL;
4408 }
4409 if (!in_bb(mem) || same_generation(mem, ld)) {
4410 #ifndef PRODUCT
4411 if (_vector_loop_debug) {
4412 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
4413 _clone_map.gen(mem->_idx));
4414 }
4415 #endif
4416 return NULL; // does not depend on loop volatile node or depends on the same generation
4417 }
4418
4419 //otherwise first node should depend on mem-phi
4420 Node* first = first_node(ld);
4421 assert(first->is_Load(), "must be Load");
4422 Node* phi = first->as_Load()->in(MemNode::Memory);
4423 if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
4424 #ifndef PRODUCT
4425 if (_vector_loop_debug) {
4426 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");
4427 ld->dump();
4428 first->dump();
4429 }
4430 #endif
4431 return NULL;
4432 }
4433
4434 Node* tail = 0;
4435 for (int m = 0; m < _mem_slice_head.length(); m++) {
4436 if (_mem_slice_head.at(m) == phi) {
4437 tail = _mem_slice_tail.at(m);
4438 }
4439 }
4440 if (tail == 0) { //test that found phi is in the list _mem_slice_head
4441 #ifndef PRODUCT
4442 if (_vector_loop_debug) {
4443 tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
4444 ld->_idx, phi->_idx);
4445 ld->dump();
4446 phi->dump();
4447 }
4448 #endif
4449 return NULL;
4450 }
4451
4452 // now all conditions are met
4453 return phi;
4454 }
4455
4456 Node* SuperWord::first_node(Node* nd) {
4457 for (int ii = 0; ii < _iteration_first.length(); ii++) {
4458 Node* nnn = _iteration_first.at(ii);
4459 if (same_origin_idx(nnn, nd)) {
4460 #ifndef PRODUCT
4461 if (_vector_loop_debug) {
4462 tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
4463 nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
4464 }
4465 #endif
4466 return nnn;
4467 }
4468 }
4469
4470 #ifndef PRODUCT
4471 if (_vector_loop_debug) {
4472 tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4473 nd->_idx, _clone_map.idx(nd->_idx));
4474 }
4475 #endif
4476 return 0;
4477 }
4478
4479 Node* SuperWord::last_node(Node* nd) {
4480 for (int ii = 0; ii < _iteration_last.length(); ii++) {
4481 Node* nnn = _iteration_last.at(ii);
4482 if (same_origin_idx(nnn, nd)) {
4483 #ifndef PRODUCT
4484 if (_vector_loop_debug) {
4485 tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4486 _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4487 }
4488 #endif
4489 return nnn;
4490 }
4491 }
4492 return 0;
4493 }
4494
4495 int SuperWord::mark_generations() {
4496 Node *ii_err = NULL, *tail_err = NULL;
4497 for (int i = 0; i < _mem_slice_head.length(); i++) {
4498 Node* phi = _mem_slice_head.at(i);
4499 assert(phi->is_Phi(), "must be phi");
4500
4501 Node* tail = _mem_slice_tail.at(i);
4502 if (_ii_last == -1) {
4503 tail_err = tail;
4504 _ii_last = _clone_map.gen(tail->_idx);
4505 }
4506 else if (_ii_last != _clone_map.gen(tail->_idx)) {
4507 #ifndef PRODUCT
4508 if (TraceSuperWord && Verbose) {
4509 tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4510 tail->dump();
4511 tail_err->dump();
4512 }
4513 #endif
4514 return -1;
4515 }
4516
4517 // find first iteration in the loop
4518 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4519 Node* ii = phi->fast_out(i);
4520 if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4521 if (_ii_first == -1) {
4522 ii_err = ii;
4523 _ii_first = _clone_map.gen(ii->_idx);
4524 } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4525 #ifndef PRODUCT
4526 if (TraceSuperWord && Verbose) {
4527 tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4528 ii->dump();
4529 if (ii_err!= 0) {
4530 ii_err->dump();
4531 }
4532 }
4533 #endif
4534 return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4535 }
4536 }
4537 }//for (DUIterator_Fast imax,
4538 }//for (int i...
4539
4540 if (_ii_first == -1 || _ii_last == -1) {
4541 if (TraceSuperWord && Verbose) {
4542 tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4543 }
4544 return -1; // something vent wrong
4545 }
4546 // collect nodes in the first and last generations
4547 assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4548 assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4549 for (int j = 0; j < _block.length(); j++) {
4550 Node* n = _block.at(j);
4551 node_idx_t gen = _clone_map.gen(n->_idx);
4552 if ((signed)gen == _ii_first) {
4553 _iteration_first.push(n);
4554 } else if ((signed)gen == _ii_last) {
4555 _iteration_last.push(n);
4556 }
4557 }
4558
4559 // building order of iterations
4560 if (_ii_order.length() == 0 && ii_err != 0) {
4561 assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4562 Node* nd = ii_err;
4563 while(_clone_map.gen(nd->_idx) != _ii_last) {
4564 _ii_order.push(_clone_map.gen(nd->_idx));
4565 bool found = false;
4566 for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4567 Node* use = nd->fast_out(i);
4568 if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4569 found = true;
4570 nd = use;
4571 break;
4572 }
4573 }//for
4574
4575 if (found == false) {
4576 if (TraceSuperWord && Verbose) {
4577 tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4578 }
4579 _ii_order.clear();
4580 return -1;
4581 }
4582 } //while
4583 _ii_order.push(_clone_map.gen(nd->_idx));
4584 }
4585
4586 #ifndef PRODUCT
4587 if (_vector_loop_debug) {
4588 tty->print_cr("SuperWord::mark_generations");
4589 tty->print_cr("First generation (%d) nodes:", _ii_first);
4590 for (int ii = 0; ii < _iteration_first.length(); ii++) _iteration_first.at(ii)->dump();
4591 tty->print_cr("Last generation (%d) nodes:", _ii_last);
4592 for (int ii = 0; ii < _iteration_last.length(); ii++) _iteration_last.at(ii)->dump();
4593 tty->print_cr(" ");
4594
4595 tty->print("SuperWord::List of generations: ");
4596 for (int jj = 0; jj < _ii_order.length(); ++jj) {
4597 tty->print("%d:%d ", jj, _ii_order.at(jj));
4598 }
4599 tty->print_cr(" ");
4600 }
4601 #endif
4602
4603 return _ii_first;
4604 }
4605
4606 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4607 assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4608 assert(same_origin_idx(gold, fix), "should be clones of the same node");
4609 Node* gin1 = gold->in(1);
4610 Node* gin2 = gold->in(2);
4611 Node* fin1 = fix->in(1);
4612 Node* fin2 = fix->in(2);
4613 bool swapped = false;
4614
4615 if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4616 if (same_origin_idx(gin1, fin1) &&
4617 same_origin_idx(gin2, fin2)) {
4618 return true; // nothing to fix
4619 }
4620 if (same_origin_idx(gin1, fin2) &&
4621 same_origin_idx(gin2, fin1)) {
4622 fix->swap_edges(1, 2);
4623 swapped = true;
4624 }
4625 }
4626 // at least one input comes from outside of bb
4627 if (gin1->_idx == fin1->_idx) {
4628 return true; // nothing to fix
4629 }
4630 if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx)) { //swapping is expensive, check condition first
4631 fix->swap_edges(1, 2);
4632 swapped = true;
4633 }
4634
4635 if (swapped) {
4636 #ifndef PRODUCT
4637 if (_vector_loop_debug) {
4638 tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4639 }
4640 #endif
4641 return true;
4642 }
4643
4644 if (TraceSuperWord && Verbose) {
4645 tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4646 }
4647
4648 return false;
4649 }
4650
4651 bool SuperWord::pack_parallel() {
4652 #ifndef PRODUCT
4653 if (_vector_loop_debug) {
4654 tty->print_cr("SuperWord::pack_parallel: START");
4655 }
4656 #endif
4657
4658 _packset.clear();
4659
4660 for (int ii = 0; ii < _iteration_first.length(); ii++) {
4661 Node* nd = _iteration_first.at(ii);
4662 if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4663 Node_List* pk = new Node_List();
4664 pk->push(nd);
4665 for (int gen = 1; gen < _ii_order.length(); ++gen) {
4666 for (int kk = 0; kk < _block.length(); kk++) {
4667 Node* clone = _block.at(kk);
4668 if (same_origin_idx(clone, nd) &&
4669 _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4670 if (nd->is_Add() || nd->is_Mul()) {
4671 fix_commutative_inputs(nd, clone);
4672 }
4673 pk->push(clone);
4674 if (pk->size() == 4) {
4675 _packset.append(pk);
4676 #ifndef PRODUCT
4677 if (_vector_loop_debug) {
4678 tty->print_cr("SuperWord::pack_parallel: added pack ");
4679 pk->dump();
4680 }
4681 #endif
4682 if (_clone_map.gen(clone->_idx) != _ii_last) {
4683 pk = new Node_List();
4684 }
4685 }
4686 break;
4687 }
4688 }
4689 }//for
4690 }//if
4691 }//for
4692
4693 #ifndef PRODUCT
4694 if (_vector_loop_debug) {
4695 tty->print_cr("SuperWord::pack_parallel: END");
4696 }
4697 #endif
4698
4699 return true;
4700 }
4701
4702 bool SuperWord::hoist_loads_in_graph() {
4703 GrowableArray<Node*> loads;
4704
4705 #ifndef PRODUCT
4706 if (_vector_loop_debug) {
4707 tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4708 }
4709 #endif
4710
4711 for (int i = 0; i < _mem_slice_head.length(); i++) {
4712 Node* n = _mem_slice_head.at(i);
4713 if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4714 if (TraceSuperWord && Verbose) {
4715 tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4716 }
4717 continue;
4718 }
4719
4720 #ifndef PRODUCT
4721 if (_vector_loop_debug) {
4722 tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d = _mem_slice_head.at(%d);", n->_idx, i);
4723 }
4724 #endif
4725
4726 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4727 Node* ld = n->fast_out(i);
4728 if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4729 for (int i = 0; i < _block.length(); i++) {
4730 Node* ld2 = _block.at(i);
4731 if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4732 !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4733 #ifndef PRODUCT
4734 if (_vector_loop_debug) {
4735 tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4736 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4737 }
4738 #endif
4739 // could not do on-the-fly, since iterator is immutable
4740 loads.push(ld2);
4741 }
4742 }// for
4743 }//if
4744 }//for (DUIterator_Fast imax,
4745 }//for (int i = 0; i
4746
4747 for (int i = 0; i < loads.length(); i++) {
4748 LoadNode* ld = loads.at(i)->as_Load();
4749 Node* phi = find_phi_for_mem_dep(ld);
4750 if (phi != NULL) {
4751 #ifndef PRODUCT
4752 if (_vector_loop_debug) {
4753 tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4754 MemNode::Memory, ld->_idx, phi->_idx);
4755 }
4756 #endif
4757 _igvn.replace_input_of(ld, MemNode::Memory, phi);
4758 }
4759 }//for
4760
4761 restart(); // invalidate all basic structures, since we rebuilt the graph
4762
4763 if (TraceSuperWord && Verbose) {
4764 tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4765 }
4766
4767 return true;
4768 }