1 /*
2 * Copyright (c) 2007, 2020, 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 if (align_to_ref() == NULL) {
503 return; // Did not find memory reference to align vectors
504 }
505
506 extend_packlist();
507
508 if (_do_vector_loop) {
509 if (_packset.length() == 0) {
510 if (TraceSuperWord) {
511 tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
512 }
513 pack_parallel();
514 }
515 }
516
517 combine_packs();
518
519 construct_my_pack_map();
520 if (UseVectorCmov) {
521 merge_packs_to_cmovd();
522 }
523
524 filter_packs();
525
526 schedule();
527 } else if (post_loop_allowed) {
528 int saved_mapped_unroll_factor = cl->slp_max_unroll();
529 if (saved_mapped_unroll_factor) {
530 int vector_mapped_unroll_factor = saved_mapped_unroll_factor;
531
532 // now reset the slp_unroll_factor so that we can check the analysis mapped
533 // what the vector loop was mapped to
534 cl->set_slp_max_unroll(0);
535
536 // do the analysis on the post loop
537 unrolling_analysis(vector_mapped_unroll_factor);
538
539 // if our analyzed loop is a canonical fit, start processing it
540 if (vector_mapped_unroll_factor == saved_mapped_unroll_factor) {
541 // now add the vector nodes to packsets
542 for (int i = 0; i < _post_block.length(); i++) {
543 Node* n = _post_block.at(i);
544 Node_List* singleton = new Node_List();
545 singleton->push(n);
546 _packset.append(singleton);
547 set_my_pack(n, singleton);
548 }
549
550 // map base types for vector usage
551 compute_vector_element_type();
552 } else {
553 return;
554 }
555 } else {
556 // for some reason we could not map the slp analysis state of the vectorized loop
557 return;
558 }
559 }
560
561 output();
562 }
563
564 //------------------------------find_adjacent_refs---------------------------
565 // Find the adjacent memory references and create pack pairs for them.
566 // This is the initial set of packs that will then be extended by
567 // following use->def and def->use links. The align positions are
568 // assigned relative to the reference "align_to_ref"
569 void SuperWord::find_adjacent_refs() {
570 // Get list of memory operations
571 Node_List memops;
572 for (int i = 0; i < _block.length(); i++) {
573 Node* n = _block.at(i);
574 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
575 is_java_primitive(n->as_Mem()->memory_type())) {
576 int align = memory_alignment(n->as_Mem(), 0);
577 if (align != bottom_align) {
578 memops.push(n);
579 }
580 }
581 }
582 if (TraceSuperWord) {
583 tty->print_cr("\nfind_adjacent_refs found %d memops", memops.size());
584 }
585
586 Node_List align_to_refs;
587 int max_idx;
588 int best_iv_adjustment = 0;
589 MemNode* best_align_to_mem_ref = NULL;
590
591 while (memops.size() != 0) {
592 // Find a memory reference to align to.
593 MemNode* mem_ref = find_align_to_ref(memops, max_idx);
594 if (mem_ref == NULL) break;
595 align_to_refs.push(mem_ref);
596 int iv_adjustment = get_iv_adjustment(mem_ref);
597
598 if (best_align_to_mem_ref == NULL) {
599 // Set memory reference which is the best from all memory operations
600 // to be used for alignment. The pre-loop trip count is modified to align
601 // this reference to a vector-aligned address.
602 best_align_to_mem_ref = mem_ref;
603 best_iv_adjustment = iv_adjustment;
604 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
605 }
606
607 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
608 // Set alignment relative to "align_to_ref" for all related memory operations.
609 for (int i = memops.size() - 1; i >= 0; i--) {
610 MemNode* s = memops.at(i)->as_Mem();
611 if (isomorphic(s, mem_ref) &&
612 (!_do_vector_loop || same_origin_idx(s, mem_ref))) {
613 SWPointer p2(s, this, NULL, false);
614 if (p2.comparable(align_to_ref_p)) {
615 int align = memory_alignment(s, iv_adjustment);
616 set_alignment(s, align);
617 }
618 }
619 }
620
621 // Create initial pack pairs of memory operations for which
622 // alignment is set and vectors will be aligned.
623 bool create_pack = true;
624 if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
625 if (vectors_should_be_aligned()) {
626 int vw = vector_width(mem_ref);
627 int vw_best = vector_width(best_align_to_mem_ref);
628 if (vw > vw_best) {
629 // Do not vectorize a memory access with more elements per vector
630 // if unaligned memory access is not allowed because number of
631 // iterations in pre-loop will be not enough to align it.
632 create_pack = false;
633 } else {
634 SWPointer p2(best_align_to_mem_ref, this, NULL, false);
635 if (align_to_ref_p.invar() != p2.invar()) {
636 // Do not vectorize memory accesses with different invariants
637 // if unaligned memory accesses are not allowed.
638 create_pack = false;
639 }
640 }
641 }
642 } else {
643 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
644 // Can't allow vectorization of unaligned memory accesses with the
645 // same type since it could be overlapped accesses to the same array.
646 create_pack = false;
647 } else {
648 // Allow independent (different type) unaligned memory operations
649 // if HW supports them.
650 if (vectors_should_be_aligned()) {
651 create_pack = false;
652 } else {
653 // Check if packs of the same memory type but
654 // with a different alignment were created before.
655 for (uint i = 0; i < align_to_refs.size(); i++) {
656 MemNode* mr = align_to_refs.at(i)->as_Mem();
657 if (mr == mem_ref) {
658 // Skip when we are looking at same memory operation.
659 continue;
660 }
661 if (same_velt_type(mr, mem_ref) &&
662 memory_alignment(mr, iv_adjustment) != 0)
663 create_pack = false;
664 }
665 }
666 }
667 }
668 if (create_pack) {
669 for (uint i = 0; i < memops.size(); i++) {
670 Node* s1 = memops.at(i);
671 int align = alignment(s1);
672 if (align == top_align) continue;
673 for (uint j = 0; j < memops.size(); j++) {
674 Node* s2 = memops.at(j);
675 if (alignment(s2) == top_align) continue;
676 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
677 if (stmts_can_pack(s1, s2, align)) {
678 Node_List* pair = new Node_List();
679 pair->push(s1);
680 pair->push(s2);
681 if (!_do_vector_loop || same_origin_idx(s1, s2)) {
682 _packset.append(pair);
683 }
684 }
685 }
686 }
687 }
688 } else { // Don't create unaligned pack
689 // First, remove remaining memory ops of the same type from the list.
690 for (int i = memops.size() - 1; i >= 0; i--) {
691 MemNode* s = memops.at(i)->as_Mem();
692 if (same_velt_type(s, mem_ref)) {
693 memops.remove(i);
694 }
695 }
696
697 // Second, remove already constructed packs of the same type.
698 for (int i = _packset.length() - 1; i >= 0; i--) {
699 Node_List* p = _packset.at(i);
700 MemNode* s = p->at(0)->as_Mem();
701 if (same_velt_type(s, mem_ref)) {
702 remove_pack_at(i);
703 }
704 }
705
706 // If needed find the best memory reference for loop alignment again.
707 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
708 // Put memory ops from remaining packs back on memops list for
709 // the best alignment search.
710 uint orig_msize = memops.size();
711 for (int i = 0; i < _packset.length(); i++) {
712 Node_List* p = _packset.at(i);
713 MemNode* s = p->at(0)->as_Mem();
714 assert(!same_velt_type(s, mem_ref), "sanity");
715 memops.push(s);
716 }
717 best_align_to_mem_ref = find_align_to_ref(memops, max_idx);
718 if (best_align_to_mem_ref == NULL) {
719 if (TraceSuperWord) {
720 tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
721 }
722 // best_align_to_mem_ref will be used for adjusting the pre-loop limit in
723 // SuperWord::align_initial_loop_index. Find one with the biggest vector size,
724 // smallest data size and smallest iv offset from memory ops from remaining packs.
725 if (_packset.length() > 0) {
726 if (orig_msize == 0) {
727 best_align_to_mem_ref = memops.at(max_idx)->as_Mem();
728 } else {
729 for (uint i = 0; i < orig_msize; i++) {
730 memops.remove(0);
731 }
732 best_align_to_mem_ref = find_align_to_ref(memops, max_idx);
733 assert(best_align_to_mem_ref == NULL, "sanity");
734 best_align_to_mem_ref = memops.at(max_idx)->as_Mem();
735 }
736 assert(best_align_to_mem_ref != NULL, "sanity");
737 }
738 break;
739 }
740 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
741 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
742 // Restore list.
743 while (memops.size() > orig_msize)
744 (void)memops.pop();
745 }
746 } // unaligned memory accesses
747
748 // Remove used mem nodes.
749 for (int i = memops.size() - 1; i >= 0; i--) {
750 MemNode* m = memops.at(i)->as_Mem();
751 if (alignment(m) != top_align) {
752 memops.remove(i);
753 }
754 }
755
756 } // while (memops.size() != 0
757 set_align_to_ref(best_align_to_mem_ref);
758
759 if (TraceSuperWord) {
760 tty->print_cr("\nAfter find_adjacent_refs");
761 print_packset();
762 }
763 }
764
765 #ifndef PRODUCT
766 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
767 if (is_trace_adjacent()) {
768 tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
769 best_align_to_mem_ref->_idx, best_iv_adjustment);
770 best_align_to_mem_ref->dump();
771 }
772 }
773 #endif
774
775 //------------------------------find_align_to_ref---------------------------
776 // Find a memory reference to align the loop induction variable to.
777 // Looks first at stores then at loads, looking for a memory reference
778 // with the largest number of references similar to it.
779 MemNode* SuperWord::find_align_to_ref(Node_List &memops, int &idx) {
780 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
781
782 // Count number of comparable memory ops
783 for (uint i = 0; i < memops.size(); i++) {
784 MemNode* s1 = memops.at(i)->as_Mem();
785 SWPointer p1(s1, this, NULL, false);
786 // Only discard unalignable memory references if vector memory references
787 // should be aligned on this platform.
788 if (vectors_should_be_aligned() && !ref_is_alignable(p1)) {
789 *cmp_ct.adr_at(i) = 0;
790 continue;
791 }
792 for (uint j = i+1; j < memops.size(); j++) {
793 MemNode* s2 = memops.at(j)->as_Mem();
794 if (isomorphic(s1, s2)) {
795 SWPointer p2(s2, this, NULL, false);
796 if (p1.comparable(p2)) {
797 (*cmp_ct.adr_at(i))++;
798 (*cmp_ct.adr_at(j))++;
799 }
800 }
801 }
802 }
803
804 // Find Store (or Load) with the greatest number of "comparable" references,
805 // biggest vector size, smallest data size and smallest iv offset.
806 int max_ct = 0;
807 int max_vw = 0;
808 int max_idx = -1;
809 int min_size = max_jint;
810 int min_iv_offset = max_jint;
811 for (uint j = 0; j < memops.size(); j++) {
812 MemNode* s = memops.at(j)->as_Mem();
813 if (s->is_Store()) {
814 int vw = vector_width_in_bytes(s);
815 assert(vw > 1, "sanity");
816 SWPointer p(s, this, NULL, false);
817 if ( cmp_ct.at(j) > max_ct ||
818 (cmp_ct.at(j) == max_ct &&
819 ( vw > max_vw ||
820 (vw == max_vw &&
821 ( data_size(s) < min_size ||
822 (data_size(s) == min_size &&
823 p.offset_in_bytes() < min_iv_offset)))))) {
824 max_ct = cmp_ct.at(j);
825 max_vw = vw;
826 max_idx = j;
827 min_size = data_size(s);
828 min_iv_offset = p.offset_in_bytes();
829 }
830 }
831 }
832 // If no stores, look at loads
833 if (max_ct == 0) {
834 for (uint j = 0; j < memops.size(); j++) {
835 MemNode* s = memops.at(j)->as_Mem();
836 if (s->is_Load()) {
837 int vw = vector_width_in_bytes(s);
838 assert(vw > 1, "sanity");
839 SWPointer p(s, this, NULL, false);
840 if ( cmp_ct.at(j) > max_ct ||
841 (cmp_ct.at(j) == max_ct &&
842 ( vw > max_vw ||
843 (vw == max_vw &&
844 ( data_size(s) < min_size ||
845 (data_size(s) == min_size &&
846 p.offset_in_bytes() < min_iv_offset)))))) {
847 max_ct = cmp_ct.at(j);
848 max_vw = vw;
849 max_idx = j;
850 min_size = data_size(s);
851 min_iv_offset = p.offset_in_bytes();
852 }
853 }
854 }
855 }
856
857 #ifdef ASSERT
858 if (TraceSuperWord && Verbose) {
859 tty->print_cr("\nVector memops after find_align_to_ref");
860 for (uint i = 0; i < memops.size(); i++) {
861 MemNode* s = memops.at(i)->as_Mem();
862 s->dump();
863 }
864 }
865 #endif
866
867 idx = max_idx;
868 if (max_ct > 0) {
869 #ifdef ASSERT
870 if (TraceSuperWord) {
871 tty->print("\nVector align to node: ");
872 memops.at(max_idx)->as_Mem()->dump();
873 }
874 #endif
875 return memops.at(max_idx)->as_Mem();
876 }
877 return NULL;
878 }
879
880 //------------------span_works_for_memory_size-----------------------------
881 static bool span_works_for_memory_size(MemNode* mem, int span, int mem_size, int offset) {
882 bool span_matches_memory = false;
883 if ((mem_size == type2aelembytes(T_BYTE) || mem_size == type2aelembytes(T_SHORT))
884 && ABS(span) == type2aelembytes(T_INT)) {
885 // There is a mismatch on span size compared to memory.
886 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
887 Node* use = mem->fast_out(j);
888 if (!VectorNode::is_type_transition_to_int(use)) {
889 return false;
890 }
891 }
892 // If all uses transition to integer, it means that we can successfully align even on mismatch.
893 return true;
894 }
895 else {
896 span_matches_memory = ABS(span) == mem_size;
897 }
898 return span_matches_memory && (ABS(offset) % mem_size) == 0;
899 }
900
901 //------------------------------ref_is_alignable---------------------------
902 // Can the preloop align the reference to position zero in the vector?
903 bool SuperWord::ref_is_alignable(SWPointer& p) {
904 if (!p.has_iv()) {
905 return true; // no induction variable
906 }
907 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
908 assert(pre_end != NULL, "we must have a correct pre-loop");
909 assert(pre_end->stride_is_con(), "pre loop stride is constant");
910 int preloop_stride = pre_end->stride_con();
911
912 int span = preloop_stride * p.scale_in_bytes();
913 int mem_size = p.memory_size();
914 int offset = p.offset_in_bytes();
915 // Stride one accesses are alignable if offset is aligned to memory operation size.
916 // Offset can be unaligned when UseUnalignedAccesses is used.
917 if (span_works_for_memory_size(p.mem(), span, mem_size, offset)) {
918 return true;
919 }
920 // If the initial offset from start of the object is computable,
921 // check if the pre-loop can align the final offset accordingly.
922 //
923 // In other words: Can we find an i such that the offset
924 // after i pre-loop iterations is aligned to vw?
925 // (init_offset + pre_loop) % vw == 0 (1)
926 // where
927 // pre_loop = i * span
928 // is the number of bytes added to the offset by i pre-loop iterations.
929 //
930 // For this to hold we need pre_loop to increase init_offset by
931 // pre_loop = vw - (init_offset % vw)
932 //
933 // This is only possible if pre_loop is divisible by span because each
934 // pre-loop iteration increases the initial offset by 'span' bytes:
935 // (vw - (init_offset % vw)) % span == 0
936 //
937 int vw = vector_width_in_bytes(p.mem());
938 assert(vw > 1, "sanity");
939 Node* init_nd = pre_end->init_trip();
940 if (init_nd->is_Con() && p.invar() == NULL) {
941 int init = init_nd->bottom_type()->is_int()->get_con();
942 int init_offset = init * p.scale_in_bytes() + offset;
943 if (init_offset < 0) { // negative offset from object start?
944 return false; // may happen in dead loop
945 }
946 if (vw % span == 0) {
947 // If vm is a multiple of span, we use formula (1).
948 if (span > 0) {
949 return (vw - (init_offset % vw)) % span == 0;
950 } else {
951 assert(span < 0, "nonzero stride * scale");
952 return (init_offset % vw) % -span == 0;
953 }
954 } else if (span % vw == 0) {
955 // If span is a multiple of vw, we can simplify formula (1) to:
956 // (init_offset + i * span) % vw == 0
957 // =>
958 // (init_offset % vw) + ((i * span) % vw) == 0
959 // =>
960 // init_offset % vw == 0
961 //
962 // Because we add a multiple of vw to the initial offset, the final
963 // offset is a multiple of vw if and only if init_offset is a multiple.
964 //
965 return (init_offset % vw) == 0;
966 }
967 }
968 return false;
969 }
970 //---------------------------get_vw_bytes_special------------------------
971 int SuperWord::get_vw_bytes_special(MemNode* s) {
972 // Get the vector width in bytes.
973 int vw = vector_width_in_bytes(s);
974
975 // Check for special case where there is an MulAddS2I usage where short vectors are going to need combined.
976 BasicType btype = velt_basic_type(s);
977 if (type2aelembytes(btype) == 2) {
978 bool should_combine_adjacent = true;
979 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
980 Node* user = s->fast_out(i);
981 if (!VectorNode::is_muladds2i(user)) {
982 should_combine_adjacent = false;
983 }
984 }
985 if (should_combine_adjacent) {
986 vw = MIN2(Matcher::max_vector_size(btype)*type2aelembytes(btype), vw * 2);
987 }
988 }
989
990 return vw;
991 }
992
993 //---------------------------get_iv_adjustment---------------------------
994 // Calculate loop's iv adjustment for this memory ops.
995 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
996 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
997 int offset = align_to_ref_p.offset_in_bytes();
998 int scale = align_to_ref_p.scale_in_bytes();
999 int elt_size = align_to_ref_p.memory_size();
1000 int vw = get_vw_bytes_special(mem_ref);
1001 assert(vw > 1, "sanity");
1002 int iv_adjustment;
1003 if (scale != 0) {
1004 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
1005 // At least one iteration is executed in pre-loop by default. As result
1006 // several iterations are needed to align memory operations in main-loop even
1007 // if offset is 0.
1008 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
1009 // iv_adjustment_in_bytes must be a multiple of elt_size if vector memory
1010 // references should be aligned on this platform.
1011 assert((ABS(iv_adjustment_in_bytes) % elt_size) == 0 || !vectors_should_be_aligned(),
1012 "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
1013 iv_adjustment = iv_adjustment_in_bytes/elt_size;
1014 } else {
1015 // This memory op is not dependent on iv (scale == 0)
1016 iv_adjustment = 0;
1017 }
1018
1019 #ifndef PRODUCT
1020 if (TraceSuperWord) {
1021 tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
1022 mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
1023 mem_ref->dump();
1024 }
1025 #endif
1026 return iv_adjustment;
1027 }
1028
1029 //---------------------------dependence_graph---------------------------
1030 // Construct dependency graph.
1031 // Add dependence edges to load/store nodes for memory dependence
1032 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
1033 void SuperWord::dependence_graph() {
1034 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
1035 // First, assign a dependence node to each memory node
1036 for (int i = 0; i < _block.length(); i++ ) {
1037 Node *n = _block.at(i);
1038 if (n->is_Mem() || (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1039 _dg.make_node(n);
1040 }
1041 }
1042
1043 // For each memory slice, create the dependences
1044 for (int i = 0; i < _mem_slice_head.length(); i++) {
1045 Node* n = _mem_slice_head.at(i);
1046 Node* n_tail = _mem_slice_tail.at(i);
1047
1048 // Get slice in predecessor order (last is first)
1049 if (cl->is_main_loop()) {
1050 mem_slice_preds(n_tail, n, _nlist);
1051 }
1052
1053 #ifndef PRODUCT
1054 if(TraceSuperWord && Verbose) {
1055 tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
1056 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
1057 _nlist.at(j)->dump();
1058 }
1059 }
1060 #endif
1061 // Make the slice dependent on the root
1062 DepMem* slice = _dg.dep(n);
1063 _dg.make_edge(_dg.root(), slice);
1064
1065 // Create a sink for the slice
1066 DepMem* slice_sink = _dg.make_node(NULL);
1067 _dg.make_edge(slice_sink, _dg.tail());
1068
1069 // Now visit each pair of memory ops, creating the edges
1070 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
1071 Node* s1 = _nlist.at(j);
1072
1073 // If no dependency yet, use slice
1074 if (_dg.dep(s1)->in_cnt() == 0) {
1075 _dg.make_edge(slice, s1);
1076 }
1077 SWPointer p1(s1->as_Mem(), this, NULL, false);
1078 bool sink_dependent = true;
1079 for (int k = j - 1; k >= 0; k--) {
1080 Node* s2 = _nlist.at(k);
1081 if (s1->is_Load() && s2->is_Load())
1082 continue;
1083 SWPointer p2(s2->as_Mem(), this, NULL, false);
1084
1085 int cmp = p1.cmp(p2);
1086 if (SuperWordRTDepCheck &&
1087 p1.base() != p2.base() && p1.valid() && p2.valid()) {
1088 // Create a runtime check to disambiguate
1089 OrderedPair pp(p1.base(), p2.base());
1090 _disjoint_ptrs.append_if_missing(pp);
1091 } else if (!SWPointer::not_equal(cmp)) {
1092 // Possibly same address
1093 _dg.make_edge(s1, s2);
1094 sink_dependent = false;
1095 }
1096 }
1097 if (sink_dependent) {
1098 _dg.make_edge(s1, slice_sink);
1099 }
1100 }
1101
1102 if (TraceSuperWord) {
1103 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
1104 for (int q = 0; q < _nlist.length(); q++) {
1105 _dg.print(_nlist.at(q));
1106 }
1107 tty->cr();
1108 }
1109
1110 _nlist.clear();
1111 }
1112
1113 if (TraceSuperWord) {
1114 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
1115 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
1116 _disjoint_ptrs.at(r).print();
1117 tty->cr();
1118 }
1119 tty->cr();
1120 }
1121
1122 }
1123
1124 //---------------------------mem_slice_preds---------------------------
1125 // Return a memory slice (node list) in predecessor order starting at "start"
1126 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
1127 assert(preds.length() == 0, "start empty");
1128 Node* n = start;
1129 Node* prev = NULL;
1130 while (true) {
1131 NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
1132 assert(in_bb(n), "must be in block");
1133 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1134 Node* out = n->fast_out(i);
1135 if (out->is_Load()) {
1136 if (in_bb(out)) {
1137 preds.push(out);
1138 if (TraceSuperWord && Verbose) {
1139 tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
1140 }
1141 }
1142 } else {
1143 // FIXME
1144 if (out->is_MergeMem() && !in_bb(out)) {
1145 // Either unrolling is causing a memory edge not to disappear,
1146 // or need to run igvn.optimize() again before SLP
1147 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
1148 // Ditto. Not sure what else to check further.
1149 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
1150 // StoreCM has an input edge used as a precedence edge.
1151 // Maybe an issue when oop stores are vectorized.
1152 } else {
1153 assert(out == prev || prev == NULL, "no branches off of store slice");
1154 }
1155 }//else
1156 }//for
1157 if (n == stop) break;
1158 preds.push(n);
1159 if (TraceSuperWord && Verbose) {
1160 tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
1161 }
1162 prev = n;
1163 assert(n->is_Mem(), "unexpected node %s", n->Name());
1164 n = n->in(MemNode::Memory);
1165 }
1166 }
1167
1168 //------------------------------stmts_can_pack---------------------------
1169 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
1170 // s1 aligned at "align"
1171 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
1172
1173 // Do not use superword for non-primitives
1174 BasicType bt1 = velt_basic_type(s1);
1175 BasicType bt2 = velt_basic_type(s2);
1176 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
1177 return false;
1178 if (Matcher::max_vector_size(bt1) < 2) {
1179 return false; // No vectors for this type
1180 }
1181
1182 if (isomorphic(s1, s2)) {
1183 if ((independent(s1, s2) && have_similar_inputs(s1, s2)) || reduction(s1, s2)) {
1184 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
1185 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
1186 int s1_align = alignment(s1);
1187 int s2_align = alignment(s2);
1188 if (s1_align == top_align || s1_align == align) {
1189 if (s2_align == top_align || s2_align == align + data_size(s1)) {
1190 return true;
1191 }
1192 }
1193 }
1194 }
1195 }
1196 }
1197 return false;
1198 }
1199
1200 //------------------------------exists_at---------------------------
1201 // Does s exist in a pack at position pos?
1202 bool SuperWord::exists_at(Node* s, uint pos) {
1203 for (int i = 0; i < _packset.length(); i++) {
1204 Node_List* p = _packset.at(i);
1205 if (p->at(pos) == s) {
1206 return true;
1207 }
1208 }
1209 return false;
1210 }
1211
1212 //------------------------------are_adjacent_refs---------------------------
1213 // Is s1 immediately before s2 in memory?
1214 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
1215 if (!s1->is_Mem() || !s2->is_Mem()) return false;
1216 if (!in_bb(s1) || !in_bb(s2)) return false;
1217
1218 // Do not use superword for non-primitives
1219 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
1220 !is_java_primitive(s2->as_Mem()->memory_type())) {
1221 return false;
1222 }
1223
1224 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
1225 // only pack memops that are in the same alias set until that's fixed.
1226 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1227 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1228 return false;
1229 SWPointer p1(s1->as_Mem(), this, NULL, false);
1230 SWPointer p2(s2->as_Mem(), this, NULL, false);
1231 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1232 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1233 return diff == data_size(s1);
1234 }
1235
1236 //------------------------------isomorphic---------------------------
1237 // Are s1 and s2 similar?
1238 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1239 if (s1->Opcode() != s2->Opcode()) return false;
1240 if (s1->req() != s2->req()) return false;
1241 if (!same_velt_type(s1, s2)) return false;
1242 Node* s1_ctrl = s1->in(0);
1243 Node* s2_ctrl = s2->in(0);
1244 // If the control nodes are equivalent, no further checks are required to test for isomorphism.
1245 if (s1_ctrl == s2_ctrl) {
1246 return true;
1247 } else {
1248 bool s1_ctrl_inv = ((s1_ctrl == NULL) ? true : lpt()->is_invariant(s1_ctrl));
1249 bool s2_ctrl_inv = ((s2_ctrl == NULL) ? true : lpt()->is_invariant(s2_ctrl));
1250 // If the control nodes are not invariant for the loop, fail isomorphism test.
1251 if (!s1_ctrl_inv || !s2_ctrl_inv) {
1252 return false;
1253 }
1254 if(s1_ctrl != NULL && s2_ctrl != NULL) {
1255 if (s1_ctrl->is_Proj()) {
1256 s1_ctrl = s1_ctrl->in(0);
1257 assert(lpt()->is_invariant(s1_ctrl), "must be invariant");
1258 }
1259 if (s2_ctrl->is_Proj()) {
1260 s2_ctrl = s2_ctrl->in(0);
1261 assert(lpt()->is_invariant(s2_ctrl), "must be invariant");
1262 }
1263 if (!s1_ctrl->is_RangeCheck() || !s2_ctrl->is_RangeCheck()) {
1264 return false;
1265 }
1266 }
1267 // Control nodes are invariant. However, we have no way of checking whether they resolve
1268 // in an equivalent manner. But, we know that invariant range checks are guaranteed to
1269 // throw before the loop (if they would have thrown). Thus, the loop would not have been reached.
1270 // Therefore, if the control nodes for both are range checks, we accept them to be isomorphic.
1271 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1272 Node* t1 = s1->fast_out(i);
1273 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1274 Node* t2 = s2->fast_out(j);
1275 if (VectorNode::is_muladds2i(t1) && VectorNode::is_muladds2i(t2)) {
1276 return true;
1277 }
1278 }
1279 }
1280 }
1281 return false;
1282 }
1283
1284 //------------------------------independent---------------------------
1285 // Is there no data path from s1 to s2 or s2 to s1?
1286 bool SuperWord::independent(Node* s1, Node* s2) {
1287 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1288 int d1 = depth(s1);
1289 int d2 = depth(s2);
1290 if (d1 == d2) return s1 != s2;
1291 Node* deep = d1 > d2 ? s1 : s2;
1292 Node* shallow = d1 > d2 ? s2 : s1;
1293
1294 visited_clear();
1295
1296 return independent_path(shallow, deep);
1297 }
1298
1299 //--------------------------have_similar_inputs-----------------------
1300 // For a node pair (s1, s2) which is isomorphic and independent,
1301 // do s1 and s2 have similar input edges?
1302 bool SuperWord::have_similar_inputs(Node* s1, Node* s2) {
1303 // assert(isomorphic(s1, s2) == true, "check isomorphic");
1304 // assert(independent(s1, s2) == true, "check independent");
1305 if (s1->req() > 1 && !s1->is_Store() && !s1->is_Load()) {
1306 for (uint i = 1; i < s1->req(); i++) {
1307 if (s1->in(i)->Opcode() != s2->in(i)->Opcode()) return false;
1308 }
1309 }
1310 return true;
1311 }
1312
1313 //------------------------------reduction---------------------------
1314 // Is there a data path between s1 and s2 and the nodes reductions?
1315 bool SuperWord::reduction(Node* s1, Node* s2) {
1316 bool retValue = false;
1317 int d1 = depth(s1);
1318 int d2 = depth(s2);
1319 if (d1 + 1 == d2) {
1320 if (s1->is_reduction() && s2->is_reduction()) {
1321 // This is an ordered set, so s1 should define s2
1322 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1323 Node* t1 = s1->fast_out(i);
1324 if (t1 == s2) {
1325 // both nodes are reductions and connected
1326 retValue = true;
1327 }
1328 }
1329 }
1330 }
1331
1332 return retValue;
1333 }
1334
1335 //------------------------------independent_path------------------------------
1336 // Helper for independent
1337 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1338 if (dp >= 1000) return false; // stop deep recursion
1339 visited_set(deep);
1340 int shal_depth = depth(shallow);
1341 assert(shal_depth <= depth(deep), "must be");
1342 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1343 Node* pred = preds.current();
1344 if (in_bb(pred) && !visited_test(pred)) {
1345 if (shallow == pred) {
1346 return false;
1347 }
1348 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1349 return false;
1350 }
1351 }
1352 }
1353 return true;
1354 }
1355
1356 //------------------------------set_alignment---------------------------
1357 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1358 set_alignment(s1, align);
1359 if (align == top_align || align == bottom_align) {
1360 set_alignment(s2, align);
1361 } else {
1362 set_alignment(s2, align + data_size(s1));
1363 }
1364 }
1365
1366 //------------------------------data_size---------------------------
1367 int SuperWord::data_size(Node* s) {
1368 Node* use = NULL; //test if the node is a candidate for CMoveV optimization, then return the size of CMov
1369 if (UseVectorCmov) {
1370 use = _cmovev_kit.is_Bool_candidate(s);
1371 if (use != NULL) {
1372 return data_size(use);
1373 }
1374 use = _cmovev_kit.is_CmpD_candidate(s);
1375 if (use != NULL) {
1376 return data_size(use);
1377 }
1378 }
1379
1380 int bsize = type2aelembytes(velt_basic_type(s));
1381 assert(bsize != 0, "valid size");
1382 return bsize;
1383 }
1384
1385 //------------------------------extend_packlist---------------------------
1386 // Extend packset by following use->def and def->use links from pack members.
1387 void SuperWord::extend_packlist() {
1388 bool changed;
1389 do {
1390 packset_sort(_packset.length());
1391 changed = false;
1392 for (int i = 0; i < _packset.length(); i++) {
1393 Node_List* p = _packset.at(i);
1394 changed |= follow_use_defs(p);
1395 changed |= follow_def_uses(p);
1396 }
1397 } while (changed);
1398
1399 if (_race_possible) {
1400 for (int i = 0; i < _packset.length(); i++) {
1401 Node_List* p = _packset.at(i);
1402 order_def_uses(p);
1403 }
1404 }
1405
1406 if (TraceSuperWord) {
1407 tty->print_cr("\nAfter extend_packlist");
1408 print_packset();
1409 }
1410 }
1411
1412 //------------------------------follow_use_defs---------------------------
1413 // Extend the packset by visiting operand definitions of nodes in pack p
1414 bool SuperWord::follow_use_defs(Node_List* p) {
1415 assert(p->size() == 2, "just checking");
1416 Node* s1 = p->at(0);
1417 Node* s2 = p->at(1);
1418 assert(s1->req() == s2->req(), "just checking");
1419 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1420
1421 if (s1->is_Load()) return false;
1422
1423 int align = alignment(s1);
1424 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1425 bool changed = false;
1426 int start = s1->is_Store() ? MemNode::ValueIn : 1;
1427 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1428 for (int j = start; j < end; j++) {
1429 Node* t1 = s1->in(j);
1430 Node* t2 = s2->in(j);
1431 if (!in_bb(t1) || !in_bb(t2))
1432 continue;
1433 if (stmts_can_pack(t1, t2, align)) {
1434 if (est_savings(t1, t2) >= 0) {
1435 Node_List* pair = new Node_List();
1436 pair->push(t1);
1437 pair->push(t2);
1438 _packset.append(pair);
1439 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1440 set_alignment(t1, t2, align);
1441 changed = true;
1442 }
1443 }
1444 }
1445 return changed;
1446 }
1447
1448 //------------------------------follow_def_uses---------------------------
1449 // Extend the packset by visiting uses of nodes in pack p
1450 bool SuperWord::follow_def_uses(Node_List* p) {
1451 bool changed = false;
1452 Node* s1 = p->at(0);
1453 Node* s2 = p->at(1);
1454 assert(p->size() == 2, "just checking");
1455 assert(s1->req() == s2->req(), "just checking");
1456 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1457
1458 if (s1->is_Store()) return false;
1459
1460 int align = alignment(s1);
1461 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1462 int savings = -1;
1463 int num_s1_uses = 0;
1464 Node* u1 = NULL;
1465 Node* u2 = NULL;
1466 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1467 Node* t1 = s1->fast_out(i);
1468 num_s1_uses++;
1469 if (!in_bb(t1)) continue;
1470 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1471 Node* t2 = s2->fast_out(j);
1472 if (!in_bb(t2)) continue;
1473 if (t2->Opcode() == Op_AddI && t2 == _lp->as_CountedLoop()->incr()) continue; // don't mess with the iv
1474 if (!opnd_positions_match(s1, t1, s2, t2))
1475 continue;
1476 if (stmts_can_pack(t1, t2, align)) {
1477 int my_savings = est_savings(t1, t2);
1478 if (my_savings > savings) {
1479 savings = my_savings;
1480 u1 = t1;
1481 u2 = t2;
1482 }
1483 }
1484 }
1485 }
1486 if (num_s1_uses > 1) {
1487 _race_possible = true;
1488 }
1489 if (savings >= 0) {
1490 Node_List* pair = new Node_List();
1491 pair->push(u1);
1492 pair->push(u2);
1493 _packset.append(pair);
1494 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1495 set_alignment(u1, u2, align);
1496 changed = true;
1497 }
1498 return changed;
1499 }
1500
1501 //------------------------------order_def_uses---------------------------
1502 // For extended packsets, ordinally arrange uses packset by major component
1503 void SuperWord::order_def_uses(Node_List* p) {
1504 Node* s1 = p->at(0);
1505
1506 if (s1->is_Store()) return;
1507
1508 // reductions are always managed beforehand
1509 if (s1->is_reduction()) return;
1510
1511 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1512 Node* t1 = s1->fast_out(i);
1513
1514 // Only allow operand swap on commuting operations
1515 if (!t1->is_Add() && !t1->is_Mul() && !VectorNode::is_muladds2i(t1)) {
1516 break;
1517 }
1518
1519 // Now find t1's packset
1520 Node_List* p2 = NULL;
1521 for (int j = 0; j < _packset.length(); j++) {
1522 p2 = _packset.at(j);
1523 Node* first = p2->at(0);
1524 if (t1 == first) {
1525 break;
1526 }
1527 p2 = NULL;
1528 }
1529 // Arrange all sub components by the major component
1530 if (p2 != NULL) {
1531 for (uint j = 1; j < p->size(); j++) {
1532 Node* d1 = p->at(j);
1533 Node* u1 = p2->at(j);
1534 opnd_positions_match(s1, t1, d1, u1);
1535 }
1536 }
1537 }
1538 }
1539
1540 //---------------------------opnd_positions_match-------------------------
1541 // Is the use of d1 in u1 at the same operand position as d2 in u2?
1542 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1543 // check reductions to see if they are marshalled to represent the reduction
1544 // operator in a specified opnd
1545 if (u1->is_reduction() && u2->is_reduction()) {
1546 // ensure reductions have phis and reduction definitions feeding the 1st operand
1547 Node* first = u1->in(2);
1548 if (first->is_Phi() || first->is_reduction()) {
1549 u1->swap_edges(1, 2);
1550 }
1551 // ensure reductions have phis and reduction definitions feeding the 1st operand
1552 first = u2->in(2);
1553 if (first->is_Phi() || first->is_reduction()) {
1554 u2->swap_edges(1, 2);
1555 }
1556 return true;
1557 }
1558
1559 uint ct = u1->req();
1560 if (ct != u2->req()) return false;
1561 uint i1 = 0;
1562 uint i2 = 0;
1563 do {
1564 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1565 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1566 if (i1 != i2) {
1567 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1568 // Further analysis relies on operands position matching.
1569 u2->swap_edges(i1, i2);
1570 } else if (VectorNode::is_muladds2i(u2) && u1 != u2) {
1571 if (i1 == 5 - i2) { // ((i1 == 3 && i2 == 2) || (i1 == 2 && i2 == 3) || (i1 == 1 && i2 == 4) || (i1 == 4 && i2 == 1))
1572 u2->swap_edges(1, 2);
1573 u2->swap_edges(3, 4);
1574 }
1575 if (i1 == 3 - i2 || i1 == 7 - i2) { // ((i1 == 1 && i2 == 2) || (i1 == 2 && i2 == 1) || (i1 == 3 && i2 == 4) || (i1 == 4 && i2 == 3))
1576 u2->swap_edges(2, 3);
1577 u2->swap_edges(1, 4);
1578 }
1579 return false; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
1580 } else {
1581 return false;
1582 }
1583 } else if (i1 == i2 && VectorNode::is_muladds2i(u2) && u1 != u2) {
1584 u2->swap_edges(1, 3);
1585 u2->swap_edges(2, 4);
1586 return false; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
1587 }
1588 } while (i1 < ct);
1589 return true;
1590 }
1591
1592 //------------------------------est_savings---------------------------
1593 // Estimate the savings from executing s1 and s2 as a pack
1594 int SuperWord::est_savings(Node* s1, Node* s2) {
1595 int save_in = 2 - 1; // 2 operations per instruction in packed form
1596
1597 // inputs
1598 for (uint i = 1; i < s1->req(); i++) {
1599 Node* x1 = s1->in(i);
1600 Node* x2 = s2->in(i);
1601 if (x1 != x2) {
1602 if (are_adjacent_refs(x1, x2)) {
1603 save_in += adjacent_profit(x1, x2);
1604 } else if (!in_packset(x1, x2)) {
1605 save_in -= pack_cost(2);
1606 } else {
1607 save_in += unpack_cost(2);
1608 }
1609 }
1610 }
1611
1612 // uses of result
1613 uint ct = 0;
1614 int save_use = 0;
1615 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1616 Node* s1_use = s1->fast_out(i);
1617 for (int j = 0; j < _packset.length(); j++) {
1618 Node_List* p = _packset.at(j);
1619 if (p->at(0) == s1_use) {
1620 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1621 Node* s2_use = s2->fast_out(k);
1622 if (p->at(p->size()-1) == s2_use) {
1623 ct++;
1624 if (are_adjacent_refs(s1_use, s2_use)) {
1625 save_use += adjacent_profit(s1_use, s2_use);
1626 }
1627 }
1628 }
1629 }
1630 }
1631 }
1632
1633 if (ct < s1->outcnt()) save_use += unpack_cost(1);
1634 if (ct < s2->outcnt()) save_use += unpack_cost(1);
1635
1636 return MAX2(save_in, save_use);
1637 }
1638
1639 //------------------------------costs---------------------------
1640 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
1641 int SuperWord::pack_cost(int ct) { return ct; }
1642 int SuperWord::unpack_cost(int ct) { return ct; }
1643
1644 //------------------------------combine_packs---------------------------
1645 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
1646 void SuperWord::combine_packs() {
1647 bool changed = true;
1648 // Combine packs regardless max vector size.
1649 while (changed) {
1650 changed = false;
1651 for (int i = 0; i < _packset.length(); i++) {
1652 Node_List* p1 = _packset.at(i);
1653 if (p1 == NULL) continue;
1654 // Because of sorting we can start at i + 1
1655 for (int j = i + 1; j < _packset.length(); j++) {
1656 Node_List* p2 = _packset.at(j);
1657 if (p2 == NULL) continue;
1658 if (i == j) continue;
1659 if (p1->at(p1->size()-1) == p2->at(0)) {
1660 for (uint k = 1; k < p2->size(); k++) {
1661 p1->push(p2->at(k));
1662 }
1663 _packset.at_put(j, NULL);
1664 changed = true;
1665 }
1666 }
1667 }
1668 }
1669
1670 // Split packs which have size greater then max vector size.
1671 for (int i = 0; i < _packset.length(); i++) {
1672 Node_List* p1 = _packset.at(i);
1673 if (p1 != NULL) {
1674 BasicType bt = velt_basic_type(p1->at(0));
1675 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1676 assert(is_power_of_2(max_vlen), "sanity");
1677 uint psize = p1->size();
1678 if (!is_power_of_2(psize)) {
1679 // Skip pack which can't be vector.
1680 // case1: for(...) { a[i] = i; } elements values are different (i+x)
1681 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
1682 _packset.at_put(i, NULL);
1683 continue;
1684 }
1685 if (psize > max_vlen) {
1686 Node_List* pack = new Node_List();
1687 for (uint j = 0; j < psize; j++) {
1688 pack->push(p1->at(j));
1689 if (pack->size() >= max_vlen) {
1690 assert(is_power_of_2(pack->size()), "sanity");
1691 _packset.append(pack);
1692 pack = new Node_List();
1693 }
1694 }
1695 _packset.at_put(i, NULL);
1696 }
1697 }
1698 }
1699
1700 // Compress list.
1701 for (int i = _packset.length() - 1; i >= 0; i--) {
1702 Node_List* p1 = _packset.at(i);
1703 if (p1 == NULL) {
1704 _packset.remove_at(i);
1705 }
1706 }
1707
1708 if (TraceSuperWord) {
1709 tty->print_cr("\nAfter combine_packs");
1710 print_packset();
1711 }
1712 }
1713
1714 //-----------------------------construct_my_pack_map--------------------------
1715 // Construct the map from nodes to packs. Only valid after the
1716 // point where a node is only in one pack (after combine_packs).
1717 void SuperWord::construct_my_pack_map() {
1718 Node_List* rslt = NULL;
1719 for (int i = 0; i < _packset.length(); i++) {
1720 Node_List* p = _packset.at(i);
1721 for (uint j = 0; j < p->size(); j++) {
1722 Node* s = p->at(j);
1723 assert(my_pack(s) == NULL, "only in one pack");
1724 set_my_pack(s, p);
1725 }
1726 }
1727 }
1728
1729 //------------------------------filter_packs---------------------------
1730 // Remove packs that are not implemented or not profitable.
1731 void SuperWord::filter_packs() {
1732 // Remove packs that are not implemented
1733 for (int i = _packset.length() - 1; i >= 0; i--) {
1734 Node_List* pk = _packset.at(i);
1735 bool impl = implemented(pk);
1736 if (!impl) {
1737 #ifndef PRODUCT
1738 if (TraceSuperWord && Verbose) {
1739 tty->print_cr("Unimplemented");
1740 pk->at(0)->dump();
1741 }
1742 #endif
1743 remove_pack_at(i);
1744 }
1745 Node *n = pk->at(0);
1746 if (n->is_reduction()) {
1747 _num_reductions++;
1748 } else {
1749 _num_work_vecs++;
1750 }
1751 }
1752
1753 // Remove packs that are not profitable
1754 bool changed;
1755 do {
1756 changed = false;
1757 for (int i = _packset.length() - 1; i >= 0; i--) {
1758 Node_List* pk = _packset.at(i);
1759 bool prof = profitable(pk);
1760 if (!prof) {
1761 #ifndef PRODUCT
1762 if (TraceSuperWord && Verbose) {
1763 tty->print_cr("Unprofitable");
1764 pk->at(0)->dump();
1765 }
1766 #endif
1767 remove_pack_at(i);
1768 changed = true;
1769 }
1770 }
1771 } while (changed);
1772
1773 #ifndef PRODUCT
1774 if (TraceSuperWord) {
1775 tty->print_cr("\nAfter filter_packs");
1776 print_packset();
1777 tty->cr();
1778 }
1779 #endif
1780 }
1781
1782 //------------------------------merge_packs_to_cmovd---------------------------
1783 // Merge CMoveD into new vector-nodes
1784 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1785 //
1786 // SubD ConD
1787 // / | /
1788 // / | / /
1789 // / | / /
1790 // / | / /
1791 // / / /
1792 // / / | /
1793 // v / | /
1794 // CmpD | /
1795 // | | /
1796 // v | /
1797 // Bool | /
1798 // \ | /
1799 // \ | /
1800 // \ | /
1801 // \ | /
1802 // \ v /
1803 // CMoveD
1804 //
1805
1806 void SuperWord::merge_packs_to_cmovd() {
1807 for (int i = _packset.length() - 1; i >= 0; i--) {
1808 _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1809 }
1810 #ifndef PRODUCT
1811 if (TraceSuperWord) {
1812 tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1813 print_packset();
1814 tty->cr();
1815 }
1816 #endif
1817 }
1818
1819 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1820 Node* use = NULL;
1821 if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1822 return NULL;
1823 }
1824 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1825 use = def->fast_out(j);
1826 if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1827 return NULL;
1828 }
1829 }
1830 return use;
1831 }
1832
1833 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1834 Node* use = NULL;
1835 if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1836 return NULL;
1837 }
1838 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1839 use = def->fast_out(j);
1840 if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1841 return NULL;
1842 }
1843 }
1844 return use;
1845 }
1846
1847 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1848 Node *cmovd = cmovd_pk->at(0);
1849 if (!cmovd->is_CMove()) {
1850 return NULL;
1851 }
1852 if (cmovd->Opcode() != Op_CMoveF && cmovd->Opcode() != Op_CMoveD) {
1853 return NULL;
1854 }
1855 if (pack(cmovd) != NULL) { // already in the cmov pack
1856 return NULL;
1857 }
1858 if (cmovd->in(0) != NULL) {
1859 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1860 return NULL;
1861 }
1862
1863 Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1864 if (!bol->is_Bool()
1865 || bol->outcnt() != 1
1866 || !_sw->same_generation(bol, cmovd)
1867 || bol->in(0) != NULL // BoolNode has control flow!!
1868 || _sw->my_pack(bol) == NULL) {
1869 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();})
1870 return NULL;
1871 }
1872 Node_List* bool_pk = _sw->my_pack(bol);
1873 if (bool_pk->size() != cmovd_pk->size() ) {
1874 return NULL;
1875 }
1876
1877 Node* cmpd = bol->in(1);
1878 if (!cmpd->is_Cmp()
1879 || cmpd->outcnt() != 1
1880 || !_sw->same_generation(cmpd, cmovd)
1881 || cmpd->in(0) != NULL // CmpDNode has control flow!!
1882 || _sw->my_pack(cmpd) == NULL) {
1883 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();})
1884 return NULL;
1885 }
1886 Node_List* cmpd_pk = _sw->my_pack(cmpd);
1887 if (cmpd_pk->size() != cmovd_pk->size() ) {
1888 return NULL;
1889 }
1890
1891 if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1892 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();})
1893 return NULL;
1894 }
1895
1896 Node_List* new_cmpd_pk = new Node_List();
1897 uint sz = cmovd_pk->size() - 1;
1898 for (uint i = 0; i <= sz; ++i) {
1899 Node* cmov = cmovd_pk->at(i);
1900 Node* bol = bool_pk->at(i);
1901 Node* cmp = cmpd_pk->at(i);
1902
1903 new_cmpd_pk->insert(i, cmov);
1904
1905 map(cmov, new_cmpd_pk);
1906 map(bol, new_cmpd_pk);
1907 map(cmp, new_cmpd_pk);
1908
1909 _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1910 }
1911 _sw->_packset.remove(cmovd_pk);
1912 _sw->_packset.remove(bool_pk);
1913 _sw->_packset.remove(cmpd_pk);
1914 _sw->_packset.append(new_cmpd_pk);
1915 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1916 return new_cmpd_pk;
1917 }
1918
1919 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1920 Node* cmpd0 = cmpd_pk->at(0);
1921 assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1922 assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1923 assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1924 Node* in1 = cmpd0->in(1);
1925 Node* in2 = cmpd0->in(2);
1926 Node_List* in1_pk = _sw->my_pack(in1);
1927 Node_List* in2_pk = _sw->my_pack(in2);
1928
1929 if ( (in1_pk != NULL && in1_pk->size() != cmpd_pk->size())
1930 || (in2_pk != NULL && in2_pk->size() != cmpd_pk->size()) ) {
1931 return false;
1932 }
1933
1934 // test if "all" in1 are in the same pack or the same node
1935 if (in1_pk == NULL) {
1936 for (uint j = 1; j < cmpd_pk->size(); j++) {
1937 if (cmpd_pk->at(j)->in(1) != in1) {
1938 return false;
1939 }
1940 }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1941 }
1942 // test if "all" in2 are in the same pack or the same node
1943 if (in2_pk == NULL) {
1944 for (uint j = 1; j < cmpd_pk->size(); j++) {
1945 if (cmpd_pk->at(j)->in(2) != in2) {
1946 return false;
1947 }
1948 }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1949 }
1950 //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1951 int cmovd_ind1, cmovd_ind2;
1952 if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1953 && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1954 cmovd_ind1 = CMoveNode::IfFalse;
1955 cmovd_ind2 = CMoveNode::IfTrue;
1956 } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1957 && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1958 cmovd_ind2 = CMoveNode::IfFalse;
1959 cmovd_ind1 = CMoveNode::IfTrue;
1960 }
1961 else {
1962 return false;
1963 }
1964
1965 for (uint j = 1; j < cmpd_pk->size(); j++) {
1966 if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1967 || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1968 return false;
1969 }//if
1970 }
1971 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(); })
1972 return true;
1973 }
1974
1975 //------------------------------implemented---------------------------
1976 // Can code be generated for pack p?
1977 bool SuperWord::implemented(Node_List* p) {
1978 bool retValue = false;
1979 Node* p0 = p->at(0);
1980 if (p0 != NULL) {
1981 int opc = p0->Opcode();
1982 uint size = p->size();
1983 if (p0->is_reduction()) {
1984 const Type *arith_type = p0->bottom_type();
1985 // Length 2 reductions of INT/LONG do not offer performance benefits
1986 if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1987 retValue = false;
1988 } else {
1989 retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1990 }
1991 } else {
1992 retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1993 }
1994 if (!retValue) {
1995 if (is_cmov_pack(p)) {
1996 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1997 return true;
1998 }
1999 }
2000 }
2001 return retValue;
2002 }
2003
2004 bool SuperWord::is_cmov_pack(Node_List* p) {
2005 return _cmovev_kit.pack(p->at(0)) != NULL;
2006 }
2007 //------------------------------same_inputs--------------------------
2008 // For pack p, are all idx operands the same?
2009 bool SuperWord::same_inputs(Node_List* p, int idx) {
2010 Node* p0 = p->at(0);
2011 uint vlen = p->size();
2012 Node* p0_def = p0->in(idx);
2013 for (uint i = 1; i < vlen; i++) {
2014 Node* pi = p->at(i);
2015 Node* pi_def = pi->in(idx);
2016 if (p0_def != pi_def) {
2017 return false;
2018 }
2019 }
2020 return true;
2021 }
2022
2023 //------------------------------profitable---------------------------
2024 // For pack p, are all operands and all uses (with in the block) vector?
2025 bool SuperWord::profitable(Node_List* p) {
2026 Node* p0 = p->at(0);
2027 uint start, end;
2028 VectorNode::vector_operands(p0, &start, &end);
2029
2030 // Return false if some inputs are not vectors or vectors with different
2031 // size or alignment.
2032 // Also, for now, return false if not scalar promotion case when inputs are
2033 // the same. Later, implement PackNode and allow differing, non-vector inputs
2034 // (maybe just the ones from outside the block.)
2035 for (uint i = start; i < end; i++) {
2036 if (!is_vector_use(p0, i)) {
2037 return false;
2038 }
2039 }
2040 // Check if reductions are connected
2041 if (p0->is_reduction()) {
2042 Node* second_in = p0->in(2);
2043 Node_List* second_pk = my_pack(second_in);
2044 if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
2045 // Remove reduction flag if no parent pack or if not enough work
2046 // to cover reduction expansion overhead
2047 p0->remove_flag(Node::Flag_is_reduction);
2048 return false;
2049 } else if (second_pk->size() != p->size()) {
2050 return false;
2051 }
2052 }
2053 if (VectorNode::is_shift(p0)) {
2054 // For now, return false if shift count is vector or not scalar promotion
2055 // case (different shift counts) because it is not supported yet.
2056 Node* cnt = p0->in(2);
2057 Node_List* cnt_pk = my_pack(cnt);
2058 if (cnt_pk != NULL)
2059 return false;
2060 if (!same_inputs(p, 2))
2061 return false;
2062 }
2063 if (!p0->is_Store()) {
2064 // For now, return false if not all uses are vector.
2065 // Later, implement ExtractNode and allow non-vector uses (maybe
2066 // just the ones outside the block.)
2067 for (uint i = 0; i < p->size(); i++) {
2068 Node* def = p->at(i);
2069 if (is_cmov_pack_internal_node(p, def)) {
2070 continue;
2071 }
2072 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2073 Node* use = def->fast_out(j);
2074 for (uint k = 0; k < use->req(); k++) {
2075 Node* n = use->in(k);
2076 if (def == n) {
2077 // Reductions should only have a Phi use at the loop head or a non-phi use
2078 // outside of the loop if it is the last element of the pack (e.g. SafePoint).
2079 if (def->is_reduction() &&
2080 ((use->is_Phi() && use->in(0) == _lpt->_head) ||
2081 (!_lpt->is_member(_phase->get_loop(_phase->ctrl_or_self(use))) && i == p->size()-1))) {
2082 continue;
2083 }
2084 if (!is_vector_use(use, k)) {
2085 return false;
2086 }
2087 }
2088 }
2089 }
2090 }
2091 }
2092 return true;
2093 }
2094
2095 //------------------------------schedule---------------------------
2096 // Adjust the memory graph for the packed operations
2097 void SuperWord::schedule() {
2098
2099 // Co-locate in the memory graph the members of each memory pack
2100 for (int i = 0; i < _packset.length(); i++) {
2101 co_locate_pack(_packset.at(i));
2102 }
2103 }
2104
2105 //-------------------------------remove_and_insert-------------------
2106 // Remove "current" from its current position in the memory graph and insert
2107 // it after the appropriate insertion point (lip or uip).
2108 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
2109 Node *uip, Unique_Node_List &sched_before) {
2110 Node* my_mem = current->in(MemNode::Memory);
2111 bool sched_up = sched_before.member(current);
2112
2113 // remove current_store from its current position in the memmory graph
2114 for (DUIterator i = current->outs(); current->has_out(i); i++) {
2115 Node* use = current->out(i);
2116 if (use->is_Mem()) {
2117 assert(use->in(MemNode::Memory) == current, "must be");
2118 if (use == prev) { // connect prev to my_mem
2119 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
2120 --i; //deleted this edge; rescan position
2121 } else if (sched_before.member(use)) {
2122 if (!sched_up) { // Will be moved together with current
2123 _igvn.replace_input_of(use, MemNode::Memory, uip);
2124 --i; //deleted this edge; rescan position
2125 }
2126 } else {
2127 if (sched_up) { // Will be moved together with current
2128 _igvn.replace_input_of(use, MemNode::Memory, lip);
2129 --i; //deleted this edge; rescan position
2130 }
2131 }
2132 }
2133 }
2134
2135 Node *insert_pt = sched_up ? uip : lip;
2136
2137 // all uses of insert_pt's memory state should use current's instead
2138 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
2139 Node* use = insert_pt->out(i);
2140 if (use->is_Mem()) {
2141 assert(use->in(MemNode::Memory) == insert_pt, "must be");
2142 _igvn.replace_input_of(use, MemNode::Memory, current);
2143 --i; //deleted this edge; rescan position
2144 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
2145 uint pos; //lip (lower insert point) must be the last one in the memory slice
2146 for (pos=1; pos < use->req(); pos++) {
2147 if (use->in(pos) == insert_pt) break;
2148 }
2149 _igvn.replace_input_of(use, pos, current);
2150 --i;
2151 }
2152 }
2153
2154 //connect current to insert_pt
2155 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
2156 }
2157
2158 //------------------------------co_locate_pack----------------------------------
2159 // To schedule a store pack, we need to move any sandwiched memory ops either before
2160 // or after the pack, based upon dependence information:
2161 // (1) If any store in the pack depends on the sandwiched memory op, the
2162 // sandwiched memory op must be scheduled BEFORE the pack;
2163 // (2) If a sandwiched memory op depends on any store in the pack, the
2164 // sandwiched memory op must be scheduled AFTER the pack;
2165 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
2166 // memory op (say memB), memB must be scheduled before memA. So, if memA is
2167 // scheduled before the pack, memB must also be scheduled before the pack;
2168 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
2169 // schedule this store AFTER the pack
2170 // (5) We know there is no dependence cycle, so there in no other case;
2171 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
2172 //
2173 // To schedule a load pack, we use the memory state of either the first or the last load in
2174 // the pack, based on the dependence constraint.
2175 void SuperWord::co_locate_pack(Node_List* pk) {
2176 if (pk->at(0)->is_Store()) {
2177 MemNode* first = executed_first(pk)->as_Mem();
2178 MemNode* last = executed_last(pk)->as_Mem();
2179 Unique_Node_List schedule_before_pack;
2180 Unique_Node_List memops;
2181
2182 MemNode* current = last->in(MemNode::Memory)->as_Mem();
2183 MemNode* previous = last;
2184 while (true) {
2185 assert(in_bb(current), "stay in block");
2186 memops.push(previous);
2187 for (DUIterator i = current->outs(); current->has_out(i); i++) {
2188 Node* use = current->out(i);
2189 if (use->is_Mem() && use != previous)
2190 memops.push(use);
2191 }
2192 if (current == first) break;
2193 previous = current;
2194 current = current->in(MemNode::Memory)->as_Mem();
2195 }
2196
2197 // determine which memory operations should be scheduled before the pack
2198 for (uint i = 1; i < memops.size(); i++) {
2199 Node *s1 = memops.at(i);
2200 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
2201 for (uint j = 0; j< i; j++) {
2202 Node *s2 = memops.at(j);
2203 if (!independent(s1, s2)) {
2204 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
2205 schedule_before_pack.push(s1); // s1 must be scheduled before
2206 Node_List* mem_pk = my_pack(s1);
2207 if (mem_pk != NULL) {
2208 for (uint ii = 0; ii < mem_pk->size(); ii++) {
2209 Node* s = mem_pk->at(ii); // follow partner
2210 if (memops.member(s) && !schedule_before_pack.member(s))
2211 schedule_before_pack.push(s);
2212 }
2213 }
2214 break;
2215 }
2216 }
2217 }
2218 }
2219 }
2220
2221 Node* upper_insert_pt = first->in(MemNode::Memory);
2222 // Following code moves loads connected to upper_insert_pt below aliased stores.
2223 // Collect such loads here and reconnect them back to upper_insert_pt later.
2224 memops.clear();
2225 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
2226 Node* use = upper_insert_pt->out(i);
2227 if (use->is_Mem() && !use->is_Store()) {
2228 memops.push(use);
2229 }
2230 }
2231
2232 MemNode* lower_insert_pt = last;
2233 previous = last; //previous store in pk
2234 current = last->in(MemNode::Memory)->as_Mem();
2235
2236 // start scheduling from "last" to "first"
2237 while (true) {
2238 assert(in_bb(current), "stay in block");
2239 assert(in_pack(previous, pk), "previous stays in pack");
2240 Node* my_mem = current->in(MemNode::Memory);
2241
2242 if (in_pack(current, pk)) {
2243 // Forward users of my memory state (except "previous) to my input memory state
2244 for (DUIterator i = current->outs(); current->has_out(i); i++) {
2245 Node* use = current->out(i);
2246 if (use->is_Mem() && use != previous) {
2247 assert(use->in(MemNode::Memory) == current, "must be");
2248 if (schedule_before_pack.member(use)) {
2249 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
2250 } else {
2251 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
2252 }
2253 --i; // deleted this edge; rescan position
2254 }
2255 }
2256 previous = current;
2257 } else { // !in_pack(current, pk) ==> a sandwiched store
2258 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
2259 }
2260
2261 if (current == first) break;
2262 current = my_mem->as_Mem();
2263 } // end while
2264
2265 // Reconnect loads back to upper_insert_pt.
2266 for (uint i = 0; i < memops.size(); i++) {
2267 Node *ld = memops.at(i);
2268 if (ld->in(MemNode::Memory) != upper_insert_pt) {
2269 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
2270 }
2271 }
2272 } else if (pk->at(0)->is_Load()) { // Load pack
2273 // All loads in the pack should have the same memory state. By default,
2274 // we use the memory state of the last load. However, if any load could
2275 // not be moved down due to the dependence constraint, we use the memory
2276 // state of the first load.
2277 Node* mem_input = pick_mem_state(pk);
2278 _igvn.hash_delete(mem_input);
2279 // Give each load the same memory state
2280 for (uint i = 0; i < pk->size(); i++) {
2281 LoadNode* ld = pk->at(i)->as_Load();
2282 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
2283 }
2284 }
2285 }
2286
2287 // Finds the first and last memory state and then picks either of them by checking dependence constraints.
2288 // If a store is dependent on an earlier load then we need to pick the memory state of the first load and cannot
2289 // pick the memory state of the last load.
2290 Node* SuperWord::pick_mem_state(Node_List* pk) {
2291 Node* first_mem = find_first_mem_state(pk);
2292 Node* last_mem = find_last_mem_state(pk, first_mem);
2293
2294 for (uint i = 0; i < pk->size(); i++) {
2295 Node* ld = pk->at(i);
2296 for (Node* current = last_mem; current != ld->in(MemNode::Memory); current = current->in(MemNode::Memory)) {
2297 assert(current->is_Mem() && in_bb(current), "unexpected memory");
2298 assert(current != first_mem, "corrupted memory graph");
2299 if (!independent(current, ld)) {
2300 #ifdef ASSERT
2301 // Added assertion code since no case has been observed that should pick the first memory state.
2302 // Remove the assertion code whenever we find a (valid) case that really needs the first memory state.
2303 pk->dump();
2304 first_mem->dump();
2305 last_mem->dump();
2306 current->dump();
2307 ld->dump();
2308 ld->in(MemNode::Memory)->dump();
2309 assert(false, "never observed that first memory should be picked");
2310 #endif
2311 return first_mem; // A later store depends on this load, pick memory state of first load
2312 }
2313 }
2314 }
2315 return last_mem;
2316 }
2317
2318 // Walk the memory graph from the current first load until the
2319 // start of the loop and check if nodes on the way are memory
2320 // edges of loads in the pack. The last one we encounter is the
2321 // first load.
2322 Node* SuperWord::find_first_mem_state(Node_List* pk) {
2323 Node* first_mem = pk->at(0)->in(MemNode::Memory);
2324 for (Node* current = first_mem; in_bb(current); current = current->is_Phi() ? current->in(LoopNode::EntryControl) : current->in(MemNode::Memory)) {
2325 assert(current->is_Mem() || (current->is_Phi() && current->in(0) == bb()), "unexpected memory");
2326 for (uint i = 1; i < pk->size(); i++) {
2327 Node* ld = pk->at(i);
2328 if (ld->in(MemNode::Memory) == current) {
2329 first_mem = current;
2330 break;
2331 }
2332 }
2333 }
2334 return first_mem;
2335 }
2336
2337 // Find the last load by going over the pack again and walking
2338 // the memory graph from the loads of the pack to the memory of
2339 // the first load. If we encounter the memory of the current last
2340 // load, then we started from further down in the memory graph and
2341 // the load we started from is the last load.
2342 Node* SuperWord::find_last_mem_state(Node_List* pk, Node* first_mem) {
2343 Node* last_mem = pk->at(0)->in(MemNode::Memory);
2344 for (uint i = 0; i < pk->size(); i++) {
2345 Node* ld = pk->at(i);
2346 for (Node* current = ld->in(MemNode::Memory); current != first_mem; current = current->in(MemNode::Memory)) {
2347 assert(current->is_Mem() && in_bb(current), "unexpected memory");
2348 if (current->in(MemNode::Memory) == last_mem) {
2349 last_mem = ld->in(MemNode::Memory);
2350 }
2351 }
2352 }
2353 return last_mem;
2354 }
2355
2356 #ifndef PRODUCT
2357 void SuperWord::print_loop(bool whole) {
2358 Node_Stack stack(_arena, _phase->C->unique() >> 2);
2359 Node_List rpo_list;
2360 VectorSet visited(_arena);
2361 visited.set(lpt()->_head->_idx);
2362 _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2363 _phase->dump(lpt(), rpo_list.size(), rpo_list );
2364 if(whole) {
2365 tty->print_cr("\n Whole loop tree");
2366 _phase->dump();
2367 tty->print_cr(" End of whole loop tree\n");
2368 }
2369 }
2370 #endif
2371
2372 //------------------------------output---------------------------
2373 // Convert packs into vector node operations
2374 void SuperWord::output() {
2375 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2376 Compile* C = _phase->C;
2377 if (_packset.length() == 0) {
2378 if (cl->is_main_loop()) {
2379 // Instigate more unrolling for optimization when vectorization fails.
2380 C->set_major_progress();
2381 cl->set_notpassed_slp();
2382 cl->mark_do_unroll_only();
2383 }
2384 return;
2385 }
2386
2387 #ifndef PRODUCT
2388 if (TraceLoopOpts) {
2389 tty->print("SuperWord::output ");
2390 lpt()->dump_head();
2391 }
2392 #endif
2393
2394 if (cl->is_main_loop()) {
2395 // MUST ENSURE main loop's initial value is properly aligned:
2396 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2397
2398 align_initial_loop_index(align_to_ref());
2399
2400 // Insert extract (unpack) operations for scalar uses
2401 for (int i = 0; i < _packset.length(); i++) {
2402 insert_extracts(_packset.at(i));
2403 }
2404 }
2405
2406 uint max_vlen_in_bytes = 0;
2407 uint max_vlen = 0;
2408 bool can_process_post_loop = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
2409
2410 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2411
2412 CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2413
2414 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2415
2416 if (do_reserve_copy() && !make_reversable.has_reserved()) {
2417 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2418 return;
2419 }
2420
2421 for (int i = 0; i < _block.length(); i++) {
2422 Node* n = _block.at(i);
2423 Node_List* p = my_pack(n);
2424 if (p && n == executed_last(p)) {
2425 uint vlen = p->size();
2426 uint vlen_in_bytes = 0;
2427 Node* vn = NULL;
2428 Node* low_adr = p->at(0);
2429 Node* first = executed_first(p);
2430 if (can_process_post_loop) {
2431 // override vlen with the main loops vector length
2432 vlen = cl->slp_max_unroll();
2433 }
2434 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);})
2435 int opc = n->Opcode();
2436 if (n->is_Load()) {
2437 Node* ctl = n->in(MemNode::Control);
2438 Node* mem = first->in(MemNode::Memory);
2439 SWPointer p1(n->as_Mem(), this, NULL, false);
2440 // Identify the memory dependency for the new loadVector node by
2441 // walking up through memory chain.
2442 // This is done to give flexibility to the new loadVector node so that
2443 // it can move above independent storeVector nodes.
2444 while (mem->is_StoreVector()) {
2445 SWPointer p2(mem->as_Mem(), this, NULL, false);
2446 int cmp = p1.cmp(p2);
2447 if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2448 mem = mem->in(MemNode::Memory);
2449 } else {
2450 break; // dependent memory
2451 }
2452 }
2453 Node* adr = low_adr->in(MemNode::Address);
2454 const TypePtr* atyp = n->adr_type();
2455 vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2456 vlen_in_bytes = vn->as_LoadVector()->memory_size();
2457 } else if (n->is_Store()) {
2458 // Promote value to be stored to vector
2459 Node* val = vector_opd(p, MemNode::ValueIn);
2460 if (val == NULL) {
2461 if (do_reserve_copy()) {
2462 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2463 return; //and reverse to backup IG
2464 }
2465 ShouldNotReachHere();
2466 }
2467
2468 Node* ctl = n->in(MemNode::Control);
2469 Node* mem = first->in(MemNode::Memory);
2470 Node* adr = low_adr->in(MemNode::Address);
2471 const TypePtr* atyp = n->adr_type();
2472 vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2473 vlen_in_bytes = vn->as_StoreVector()->memory_size();
2474 } else if (VectorNode::is_scalar_rotate(n)) {
2475 Node* in1 = low_adr->in(1);
2476 Node* in2 = p->at(0)->in(2);
2477 assert(in2->bottom_type()->isa_int(), "Shift must always be an int value");
2478 // If rotation count is non-constant or greater than 8bit value create a vector.
2479 if (!in2->is_Con() || -0x80 > in2->get_int() || in2->get_int() >= 0x80) {
2480 in2 = vector_opd(p, 2);
2481 }
2482 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2483 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2484 } else if (VectorNode::is_roundopD(n)) {
2485 Node* in1 = vector_opd(p, 1);
2486 Node* in2 = low_adr->in(2);
2487 assert(in2->is_Con(), "Constant rounding mode expected.");
2488 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2489 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2490 } else if (VectorNode::is_muladds2i(n)) {
2491 assert(n->req() == 5u, "MulAddS2I should have 4 operands.");
2492 Node* in1 = vector_opd(p, 1);
2493 Node* in2 = vector_opd(p, 2);
2494 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2495 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2496 } else if (n->req() == 3 && !is_cmov_pack(p)) {
2497 // Promote operands to vector
2498 Node* in1 = NULL;
2499 bool node_isa_reduction = n->is_reduction();
2500 if (node_isa_reduction) {
2501 // the input to the first reduction operation is retained
2502 in1 = low_adr->in(1);
2503 } else {
2504 in1 = vector_opd(p, 1);
2505 if (in1 == NULL) {
2506 if (do_reserve_copy()) {
2507 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2508 return; //and reverse to backup IG
2509 }
2510 ShouldNotReachHere();
2511 }
2512 }
2513 Node* in2 = vector_opd(p, 2);
2514 if (in2 == NULL) {
2515 if (do_reserve_copy()) {
2516 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2517 return; //and reverse to backup IG
2518 }
2519 ShouldNotReachHere();
2520 }
2521 if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2522 // Move invariant vector input into second position to avoid register spilling.
2523 Node* tmp = in1;
2524 in1 = in2;
2525 in2 = tmp;
2526 }
2527 if (node_isa_reduction) {
2528 const Type *arith_type = n->bottom_type();
2529 vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2530 if (in2->is_Load()) {
2531 vlen_in_bytes = in2->as_LoadVector()->memory_size();
2532 } else {
2533 vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2534 }
2535 } else {
2536 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2537 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2538 }
2539 } else if (opc == Op_SqrtF || opc == Op_SqrtD ||
2540 opc == Op_AbsF || opc == Op_AbsD ||
2541 opc == Op_AbsI || opc == Op_AbsL ||
2542 opc == Op_NegF || opc == Op_NegD ||
2543 opc == Op_PopCountI) {
2544 assert(n->req() == 2, "only one input expected");
2545 Node* in = vector_opd(p, 1);
2546 vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2547 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2548 } else if (is_cmov_pack(p)) {
2549 if (can_process_post_loop) {
2550 // do not refactor of flow in post loop context
2551 return;
2552 }
2553 if (!n->is_CMove()) {
2554 continue;
2555 }
2556 // place here CMoveVDNode
2557 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2558 Node* bol = n->in(CMoveNode::Condition);
2559 if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2560 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();})
2561 bol = bol->in(1); //may be ExtractNode
2562 }
2563
2564 assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2565 if (!bol->is_Bool()) {
2566 if (do_reserve_copy()) {
2567 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2568 return; //and reverse to backup IG
2569 }
2570 ShouldNotReachHere();
2571 }
2572
2573 int cond = (int)bol->as_Bool()->_test._test;
2574 Node* in_cc = _igvn.intcon(cond);
2575 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2576 Node* cc = bol->clone();
2577 cc->set_req(1, in_cc);
2578 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2579
2580 Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2581 if (src1 == NULL) {
2582 if (do_reserve_copy()) {
2583 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2584 return; //and reverse to backup IG
2585 }
2586 ShouldNotReachHere();
2587 }
2588 Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2589 if (src2 == NULL) {
2590 if (do_reserve_copy()) {
2591 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2592 return; //and reverse to backup IG
2593 }
2594 ShouldNotReachHere();
2595 }
2596 BasicType bt = velt_basic_type(n);
2597 const TypeVect* vt = TypeVect::make(bt, vlen);
2598 assert(bt == T_FLOAT || bt == T_DOUBLE, "Only vectorization for FP cmovs is supported");
2599 if (bt == T_FLOAT) {
2600 vn = new CMoveVFNode(cc, src1, src2, vt);
2601 } else {
2602 assert(bt == T_DOUBLE, "Expected double");
2603 vn = new CMoveVDNode(cc, src1, src2, vt);
2604 }
2605 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2606 } else if (opc == Op_FmaD || opc == Op_FmaF) {
2607 // Promote operands to vector
2608 Node* in1 = vector_opd(p, 1);
2609 Node* in2 = vector_opd(p, 2);
2610 Node* in3 = vector_opd(p, 3);
2611 vn = VectorNode::make(opc, in1, in2, in3, vlen, velt_basic_type(n));
2612 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2613 } else {
2614 if (do_reserve_copy()) {
2615 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2616 return; //and reverse to backup IG
2617 }
2618 ShouldNotReachHere();
2619 }
2620
2621 assert(vn != NULL, "sanity");
2622 if (vn == NULL) {
2623 if (do_reserve_copy()){
2624 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2625 return; //and reverse to backup IG
2626 }
2627 ShouldNotReachHere();
2628 }
2629
2630 _block.at_put(i, vn);
2631 _igvn.register_new_node_with_optimizer(vn);
2632 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2633 for (uint j = 0; j < p->size(); j++) {
2634 Node* pm = p->at(j);
2635 _igvn.replace_node(pm, vn);
2636 }
2637 _igvn._worklist.push(vn);
2638
2639 if (can_process_post_loop) {
2640 // first check if the vector size if the maximum vector which we can use on the machine,
2641 // other vector size have reduced values for predicated data mapping.
2642 if (vlen_in_bytes != (uint)MaxVectorSize) {
2643 return;
2644 }
2645 }
2646
2647 if (vlen > max_vlen) {
2648 max_vlen = vlen;
2649 }
2650 if (vlen_in_bytes > max_vlen_in_bytes) {
2651 max_vlen_in_bytes = vlen_in_bytes;
2652 }
2653 #ifdef ASSERT
2654 if (TraceNewVectors) {
2655 tty->print("new Vector node: ");
2656 vn->dump();
2657 }
2658 #endif
2659 }
2660 }//for (int i = 0; i < _block.length(); i++)
2661
2662 if (max_vlen_in_bytes > C->max_vector_size()) {
2663 C->set_max_vector_size(max_vlen_in_bytes);
2664 }
2665 if (max_vlen_in_bytes > 0) {
2666 cl->mark_loop_vectorized();
2667 }
2668
2669 if (SuperWordLoopUnrollAnalysis) {
2670 if (cl->has_passed_slp()) {
2671 uint slp_max_unroll_factor = cl->slp_max_unroll();
2672 if (slp_max_unroll_factor == max_vlen) {
2673 if (TraceSuperWordLoopUnrollAnalysis) {
2674 tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2675 }
2676
2677 // For atomic unrolled loops which are vector mapped, instigate more unrolling
2678 cl->set_notpassed_slp();
2679 if (cl->is_main_loop()) {
2680 // if vector resources are limited, do not allow additional unrolling, also
2681 // do not unroll more on pure vector loops which were not reduced so that we can
2682 // program the post loop to single iteration execution.
2683 if (FLOATPRESSURE > 8) {
2684 C->set_major_progress();
2685 cl->mark_do_unroll_only();
2686 }
2687 }
2688
2689 if (do_reserve_copy()) {
2690 if (can_process_post_loop) {
2691 // Now create the difference of trip and limit and use it as our mask index.
2692 // Note: We limited the unroll of the vectorized loop so that
2693 // only vlen-1 size iterations can remain to be mask programmed.
2694 Node *incr = cl->incr();
2695 SubINode *index = new SubINode(cl->limit(), cl->init_trip());
2696 _igvn.register_new_node_with_optimizer(index);
2697 SetVectMaskINode *mask = new SetVectMaskINode(_phase->get_ctrl(cl->init_trip()), index);
2698 _igvn.register_new_node_with_optimizer(mask);
2699 // make this a single iteration loop
2700 AddINode *new_incr = new AddINode(incr->in(1), mask);
2701 _igvn.register_new_node_with_optimizer(new_incr);
2702 _phase->set_ctrl(new_incr, _phase->get_ctrl(incr));
2703 _igvn.replace_node(incr, new_incr);
2704 cl->mark_is_multiversioned();
2705 cl->loopexit()->add_flag(Node::Flag_has_vector_mask_set);
2706 }
2707 }
2708 }
2709 }
2710 }
2711
2712 if (do_reserve_copy()) {
2713 make_reversable.use_new();
2714 }
2715 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2716 return;
2717 }
2718
2719 //------------------------------vector_opd---------------------------
2720 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
2721 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2722 Node* p0 = p->at(0);
2723 uint vlen = p->size();
2724 Node* opd = p0->in(opd_idx);
2725 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2726
2727 if (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()) {
2728 // override vlen with the main loops vector length
2729 vlen = cl->slp_max_unroll();
2730 }
2731
2732 if (same_inputs(p, opd_idx)) {
2733 if (opd->is_Vector() || opd->is_LoadVector()) {
2734 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2735 if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2736 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2737 return NULL;
2738 }
2739 return opd; // input is matching vector
2740 }
2741 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2742 Compile* C = _phase->C;
2743 Node* cnt = opd;
2744 // Vector instructions do not mask shift count, do it here.
2745 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2746 const TypeInt* t = opd->find_int_type();
2747 if (t != NULL && t->is_con()) {
2748 juint shift = t->get_con();
2749 if (shift > mask) { // Unsigned cmp
2750 cnt = ConNode::make(TypeInt::make(shift & mask));
2751 }
2752 } else {
2753 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2754 cnt = ConNode::make(TypeInt::make(mask));
2755 _igvn.register_new_node_with_optimizer(cnt);
2756 cnt = new AndINode(opd, cnt);
2757 _igvn.register_new_node_with_optimizer(cnt);
2758 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2759 }
2760 assert(opd->bottom_type()->isa_int(), "int type only");
2761 if (!opd->bottom_type()->isa_int()) {
2762 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2763 return NULL;
2764 }
2765 }
2766 // Move shift count into vector register.
2767 cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2768 _igvn.register_new_node_with_optimizer(cnt);
2769 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2770 return cnt;
2771 }
2772 assert(!opd->is_StoreVector(), "such vector is not expected here");
2773 if (opd->is_StoreVector()) {
2774 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2775 return NULL;
2776 }
2777 // Convert scalar input to vector with the same number of elements as
2778 // p0's vector. Use p0's type because size of operand's container in
2779 // vector should match p0's size regardless operand's size.
2780 const Type* p0_t = NULL;
2781 VectorNode* vn = NULL;
2782 if (opd_idx == 2 && VectorNode::is_scalar_rotate(p0)) {
2783 Node* conv = opd;
2784 p0_t = TypeInt::INT;
2785 if (p0->bottom_type()->isa_long()) {
2786 p0_t = TypeLong::LONG;
2787 conv = new ConvI2LNode(opd);
2788 _igvn.register_new_node_with_optimizer(conv);
2789 _phase->set_ctrl(conv, _phase->get_ctrl(opd));
2790 }
2791 vn = VectorNode::scalar2vector(conv, vlen, p0_t);
2792 } else {
2793 p0_t = velt_type(p0);
2794 vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2795 }
2796
2797 _igvn.register_new_node_with_optimizer(vn);
2798 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2799 #ifdef ASSERT
2800 if (TraceNewVectors) {
2801 tty->print("new Vector node: ");
2802 vn->dump();
2803 }
2804 #endif
2805 return vn;
2806 }
2807
2808 // Insert pack operation
2809 BasicType bt = velt_basic_type(p0);
2810 PackNode* pk = PackNode::make(opd, vlen, bt);
2811 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2812
2813 for (uint i = 1; i < vlen; i++) {
2814 Node* pi = p->at(i);
2815 Node* in = pi->in(opd_idx);
2816 assert(my_pack(in) == NULL, "Should already have been unpacked");
2817 if (my_pack(in) != NULL) {
2818 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2819 return NULL;
2820 }
2821 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2822 pk->add_opd(in);
2823 if (VectorNode::is_muladds2i(pi)) {
2824 Node* in2 = pi->in(opd_idx + 2);
2825 assert(my_pack(in2) == NULL, "Should already have been unpacked");
2826 if (my_pack(in2) != NULL) {
2827 NOT_PRODUCT(if (is_trace_loop_reverse() || TraceLoopOpts) { tty->print_cr("Should already have been unpacked"); })
2828 return NULL;
2829 }
2830 assert(opd_bt == in2->bottom_type()->basic_type(), "all same type");
2831 pk->add_opd(in2);
2832 }
2833 }
2834 _igvn.register_new_node_with_optimizer(pk);
2835 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2836 #ifdef ASSERT
2837 if (TraceNewVectors) {
2838 tty->print("new Vector node: ");
2839 pk->dump();
2840 }
2841 #endif
2842 return pk;
2843 }
2844
2845 //------------------------------insert_extracts---------------------------
2846 // If a use of pack p is not a vector use, then replace the
2847 // use with an extract operation.
2848 void SuperWord::insert_extracts(Node_List* p) {
2849 if (p->at(0)->is_Store()) return;
2850 assert(_n_idx_list.is_empty(), "empty (node,index) list");
2851
2852 // Inspect each use of each pack member. For each use that is
2853 // not a vector use, replace the use with an extract operation.
2854
2855 for (uint i = 0; i < p->size(); i++) {
2856 Node* def = p->at(i);
2857 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2858 Node* use = def->fast_out(j);
2859 for (uint k = 0; k < use->req(); k++) {
2860 Node* n = use->in(k);
2861 if (def == n) {
2862 Node_List* u_pk = my_pack(use);
2863 if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2864 _n_idx_list.push(use, k);
2865 }
2866 }
2867 }
2868 }
2869 }
2870
2871 while (_n_idx_list.is_nonempty()) {
2872 Node* use = _n_idx_list.node();
2873 int idx = _n_idx_list.index();
2874 _n_idx_list.pop();
2875 Node* def = use->in(idx);
2876
2877 if (def->is_reduction()) continue;
2878
2879 // Insert extract operation
2880 _igvn.hash_delete(def);
2881 int def_pos = alignment(def) / data_size(def);
2882
2883 Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2884 _igvn.register_new_node_with_optimizer(ex);
2885 _phase->set_ctrl(ex, _phase->get_ctrl(def));
2886 _igvn.replace_input_of(use, idx, ex);
2887 _igvn._worklist.push(def);
2888
2889 bb_insert_after(ex, bb_idx(def));
2890 set_velt_type(ex, velt_type(def));
2891 }
2892 }
2893
2894 //------------------------------is_vector_use---------------------------
2895 // Is use->in(u_idx) a vector use?
2896 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2897 Node_List* u_pk = my_pack(use);
2898 if (u_pk == NULL) return false;
2899 if (use->is_reduction()) return true;
2900 Node* def = use->in(u_idx);
2901 Node_List* d_pk = my_pack(def);
2902 if (d_pk == NULL) {
2903 // check for scalar promotion
2904 Node* n = u_pk->at(0)->in(u_idx);
2905 for (uint i = 1; i < u_pk->size(); i++) {
2906 if (u_pk->at(i)->in(u_idx) != n) return false;
2907 }
2908 return true;
2909 }
2910 if (VectorNode::is_muladds2i(use)) {
2911 // MulAddS2I takes shorts and produces ints - hence the special checks
2912 // on alignment and size.
2913 if (u_pk->size() * 2 != d_pk->size()) {
2914 return false;
2915 }
2916 for (uint i = 0; i < MIN2(d_pk->size(), u_pk->size()); i++) {
2917 Node* ui = u_pk->at(i);
2918 Node* di = d_pk->at(i);
2919 if (alignment(ui) != alignment(di) * 2) {
2920 return false;
2921 }
2922 }
2923 return true;
2924 }
2925 if (u_pk->size() != d_pk->size())
2926 return false;
2927 for (uint i = 0; i < u_pk->size(); i++) {
2928 Node* ui = u_pk->at(i);
2929 Node* di = d_pk->at(i);
2930 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2931 return false;
2932 }
2933 return true;
2934 }
2935
2936 //------------------------------construct_bb---------------------------
2937 // Construct reverse postorder list of block members
2938 bool SuperWord::construct_bb() {
2939 Node* entry = bb();
2940
2941 assert(_stk.length() == 0, "stk is empty");
2942 assert(_block.length() == 0, "block is empty");
2943 assert(_data_entry.length() == 0, "data_entry is empty");
2944 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2945 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2946
2947 // Find non-control nodes with no inputs from within block,
2948 // create a temporary map from node _idx to bb_idx for use
2949 // by the visited and post_visited sets,
2950 // and count number of nodes in block.
2951 int bb_ct = 0;
2952 for (uint i = 0; i < lpt()->_body.size(); i++) {
2953 Node *n = lpt()->_body.at(i);
2954 set_bb_idx(n, i); // Create a temporary map
2955 if (in_bb(n)) {
2956 if (n->is_LoadStore() || n->is_MergeMem() ||
2957 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2958 // Bailout if the loop has LoadStore, MergeMem or data Proj
2959 // nodes. Superword optimization does not work with them.
2960 return false;
2961 }
2962 bb_ct++;
2963 if (!n->is_CFG()) {
2964 bool found = false;
2965 for (uint j = 0; j < n->req(); j++) {
2966 Node* def = n->in(j);
2967 if (def && in_bb(def)) {
2968 found = true;
2969 break;
2970 }
2971 }
2972 if (!found) {
2973 assert(n != entry, "can't be entry");
2974 _data_entry.push(n);
2975 }
2976 }
2977 }
2978 }
2979
2980 // Find memory slices (head and tail)
2981 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2982 Node *n = lp()->fast_out(i);
2983 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2984 Node* n_tail = n->in(LoopNode::LoopBackControl);
2985 if (n_tail != n->in(LoopNode::EntryControl)) {
2986 if (!n_tail->is_Mem()) {
2987 assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2988 return false; // Bailout
2989 }
2990 _mem_slice_head.push(n);
2991 _mem_slice_tail.push(n_tail);
2992 }
2993 }
2994 }
2995
2996 // Create an RPO list of nodes in block
2997
2998 visited_clear();
2999 post_visited_clear();
3000
3001 // Push all non-control nodes with no inputs from within block, then control entry
3002 for (int j = 0; j < _data_entry.length(); j++) {
3003 Node* n = _data_entry.at(j);
3004 visited_set(n);
3005 _stk.push(n);
3006 }
3007 visited_set(entry);
3008 _stk.push(entry);
3009
3010 // Do a depth first walk over out edges
3011 int rpo_idx = bb_ct - 1;
3012 int size;
3013 int reduction_uses = 0;
3014 while ((size = _stk.length()) > 0) {
3015 Node* n = _stk.top(); // Leave node on stack
3016 if (!visited_test_set(n)) {
3017 // forward arc in graph
3018 } else if (!post_visited_test(n)) {
3019 // cross or back arc
3020 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3021 Node *use = n->fast_out(i);
3022 if (in_bb(use) && !visited_test(use) &&
3023 // Don't go around backedge
3024 (!use->is_Phi() || n == entry)) {
3025 if (use->is_reduction()) {
3026 // First see if we can map the reduction on the given system we are on, then
3027 // make a data entry operation for each reduction we see.
3028 BasicType bt = use->bottom_type()->basic_type();
3029 if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
3030 reduction_uses++;
3031 }
3032 }
3033 _stk.push(use);
3034 }
3035 }
3036 if (_stk.length() == size) {
3037 // There were no additional uses, post visit node now
3038 _stk.pop(); // Remove node from stack
3039 assert(rpo_idx >= 0, "");
3040 _block.at_put_grow(rpo_idx, n);
3041 rpo_idx--;
3042 post_visited_set(n);
3043 assert(rpo_idx >= 0 || _stk.is_empty(), "");
3044 }
3045 } else {
3046 _stk.pop(); // Remove post-visited node from stack
3047 }
3048 }//while
3049
3050 int ii_current = -1;
3051 unsigned int load_idx = (unsigned int)-1;
3052 _ii_order.clear();
3053 // Create real map of block indices for nodes
3054 for (int j = 0; j < _block.length(); j++) {
3055 Node* n = _block.at(j);
3056 set_bb_idx(n, j);
3057 if (_do_vector_loop && n->is_Load()) {
3058 if (ii_current == -1) {
3059 ii_current = _clone_map.gen(n->_idx);
3060 _ii_order.push(ii_current);
3061 load_idx = _clone_map.idx(n->_idx);
3062 } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
3063 ii_current = _clone_map.gen(n->_idx);
3064 _ii_order.push(ii_current);
3065 }
3066 }
3067 }//for
3068
3069 // Ensure extra info is allocated.
3070 initialize_bb();
3071
3072 #ifndef PRODUCT
3073 if (_vector_loop_debug && _ii_order.length() > 0) {
3074 tty->print("SuperWord::construct_bb: List of generations: ");
3075 for (int jj = 0; jj < _ii_order.length(); ++jj) {
3076 tty->print(" %d:%d", jj, _ii_order.at(jj));
3077 }
3078 tty->print_cr(" ");
3079 }
3080 if (TraceSuperWord) {
3081 print_bb();
3082 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
3083 for (int m = 0; m < _data_entry.length(); m++) {
3084 tty->print("%3d ", m);
3085 _data_entry.at(m)->dump();
3086 }
3087 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
3088 for (int m = 0; m < _mem_slice_head.length(); m++) {
3089 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
3090 tty->print(" "); _mem_slice_tail.at(m)->dump();
3091 }
3092 }
3093 #endif
3094 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
3095 return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
3096 }
3097
3098 //------------------------------initialize_bb---------------------------
3099 // Initialize per node info
3100 void SuperWord::initialize_bb() {
3101 Node* last = _block.at(_block.length() - 1);
3102 grow_node_info(bb_idx(last));
3103 }
3104
3105 //------------------------------bb_insert_after---------------------------
3106 // Insert n into block after pos
3107 void SuperWord::bb_insert_after(Node* n, int pos) {
3108 int n_pos = pos + 1;
3109 // Make room
3110 for (int i = _block.length() - 1; i >= n_pos; i--) {
3111 _block.at_put_grow(i+1, _block.at(i));
3112 }
3113 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
3114 _node_info.at_put_grow(j+1, _node_info.at(j));
3115 }
3116 // Set value
3117 _block.at_put_grow(n_pos, n);
3118 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
3119 // Adjust map from node->_idx to _block index
3120 for (int i = n_pos; i < _block.length(); i++) {
3121 set_bb_idx(_block.at(i), i);
3122 }
3123 }
3124
3125 //------------------------------compute_max_depth---------------------------
3126 // Compute max depth for expressions from beginning of block
3127 // Use to prune search paths during test for independence.
3128 void SuperWord::compute_max_depth() {
3129 int ct = 0;
3130 bool again;
3131 do {
3132 again = false;
3133 for (int i = 0; i < _block.length(); i++) {
3134 Node* n = _block.at(i);
3135 if (!n->is_Phi()) {
3136 int d_orig = depth(n);
3137 int d_in = 0;
3138 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
3139 Node* pred = preds.current();
3140 if (in_bb(pred)) {
3141 d_in = MAX2(d_in, depth(pred));
3142 }
3143 }
3144 if (d_in + 1 != d_orig) {
3145 set_depth(n, d_in + 1);
3146 again = true;
3147 }
3148 }
3149 }
3150 ct++;
3151 } while (again);
3152
3153 if (TraceSuperWord && Verbose) {
3154 tty->print_cr("compute_max_depth iterated: %d times", ct);
3155 }
3156 }
3157
3158 //-------------------------compute_vector_element_type-----------------------
3159 // Compute necessary vector element type for expressions
3160 // This propagates backwards a narrower integer type when the
3161 // upper bits of the value are not needed.
3162 // Example: char a,b,c; a = b + c;
3163 // Normally the type of the add is integer, but for packed character
3164 // operations the type of the add needs to be char.
3165 void SuperWord::compute_vector_element_type() {
3166 if (TraceSuperWord && Verbose) {
3167 tty->print_cr("\ncompute_velt_type:");
3168 }
3169
3170 // Initial type
3171 for (int i = 0; i < _block.length(); i++) {
3172 Node* n = _block.at(i);
3173 set_velt_type(n, container_type(n));
3174 }
3175
3176 // Propagate integer narrowed type backwards through operations
3177 // that don't depend on higher order bits
3178 for (int i = _block.length() - 1; i >= 0; i--) {
3179 Node* n = _block.at(i);
3180 // Only integer types need be examined
3181 const Type* vtn = velt_type(n);
3182 if (vtn->basic_type() == T_INT) {
3183 uint start, end;
3184 VectorNode::vector_operands(n, &start, &end);
3185
3186 for (uint j = start; j < end; j++) {
3187 Node* in = n->in(j);
3188 // Don't propagate through a memory
3189 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
3190 data_size(n) < data_size(in)) {
3191 bool same_type = true;
3192 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
3193 Node *use = in->fast_out(k);
3194 if (!in_bb(use) || !same_velt_type(use, n)) {
3195 same_type = false;
3196 break;
3197 }
3198 }
3199 if (same_type) {
3200 // For right shifts of small integer types (bool, byte, char, short)
3201 // we need precise information about sign-ness. Only Load nodes have
3202 // this information because Store nodes are the same for signed and
3203 // unsigned values. And any arithmetic operation after a load may
3204 // expand a value to signed Int so such right shifts can't be used
3205 // because vector elements do not have upper bits of Int.
3206 const Type* vt = vtn;
3207 if (VectorNode::is_shift(in)) {
3208 Node* load = in->in(1);
3209 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
3210 vt = velt_type(load);
3211 } else if (in->Opcode() != Op_LShiftI) {
3212 // Widen type to Int to avoid creation of right shift vector
3213 // (align + data_size(s1) check in stmts_can_pack() will fail).
3214 // Note, left shifts work regardless type.
3215 vt = TypeInt::INT;
3216 }
3217 }
3218 set_velt_type(in, vt);
3219 }
3220 }
3221 }
3222 }
3223 }
3224 #ifndef PRODUCT
3225 if (TraceSuperWord && Verbose) {
3226 for (int i = 0; i < _block.length(); i++) {
3227 Node* n = _block.at(i);
3228 velt_type(n)->dump();
3229 tty->print("\t");
3230 n->dump();
3231 }
3232 }
3233 #endif
3234 }
3235
3236 //------------------------------memory_alignment---------------------------
3237 // Alignment within a vector memory reference
3238 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
3239 #ifndef PRODUCT
3240 if ((TraceSuperWord && Verbose) || is_trace_alignment()) {
3241 tty->print("SuperWord::memory_alignment within a vector memory reference for %d: ", s->_idx); s->dump();
3242 }
3243 #endif
3244 NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
3245 SWPointer p(s, this, NULL, false);
3246 if (!p.valid()) {
3247 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
3248 return bottom_align;
3249 }
3250 int vw = get_vw_bytes_special(s);
3251 if (vw < 2) {
3252 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
3253 return bottom_align; // No vectors for this type
3254 }
3255 int offset = p.offset_in_bytes();
3256 offset += iv_adjust*p.memory_size();
3257 int off_rem = offset % vw;
3258 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
3259 #ifndef PRODUCT
3260 if ((TraceSuperWord && Verbose) || is_trace_alignment()) {
3261 tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
3262 }
3263 #endif
3264 return off_mod;
3265 }
3266
3267 //---------------------------container_type---------------------------
3268 // Smallest type containing range of values
3269 const Type* SuperWord::container_type(Node* n) {
3270 if (n->is_Mem()) {
3271 BasicType bt = n->as_Mem()->memory_type();
3272 if (n->is_Store() && (bt == T_CHAR)) {
3273 // Use T_SHORT type instead of T_CHAR for stored values because any
3274 // preceding arithmetic operation extends values to signed Int.
3275 bt = T_SHORT;
3276 }
3277 if (n->Opcode() == Op_LoadUB) {
3278 // Adjust type for unsigned byte loads, it is important for right shifts.
3279 // T_BOOLEAN is used because there is no basic type representing type
3280 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
3281 // size (one byte) and sign is important.
3282 bt = T_BOOLEAN;
3283 }
3284 return Type::get_const_basic_type(bt);
3285 }
3286 const Type* t = _igvn.type(n);
3287 if (t->basic_type() == T_INT) {
3288 // A narrow type of arithmetic operations will be determined by
3289 // propagating the type of memory operations.
3290 return TypeInt::INT;
3291 }
3292 return t;
3293 }
3294
3295 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
3296 const Type* vt1 = velt_type(n1);
3297 const Type* vt2 = velt_type(n2);
3298 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
3299 // Compare vectors element sizes for integer types.
3300 return data_size(n1) == data_size(n2);
3301 }
3302 return vt1 == vt2;
3303 }
3304
3305 //------------------------------in_packset---------------------------
3306 // Are s1 and s2 in a pack pair and ordered as s1,s2?
3307 bool SuperWord::in_packset(Node* s1, Node* s2) {
3308 for (int i = 0; i < _packset.length(); i++) {
3309 Node_List* p = _packset.at(i);
3310 assert(p->size() == 2, "must be");
3311 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
3312 return true;
3313 }
3314 }
3315 return false;
3316 }
3317
3318 //------------------------------in_pack---------------------------
3319 // Is s in pack p?
3320 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
3321 for (uint i = 0; i < p->size(); i++) {
3322 if (p->at(i) == s) {
3323 return p;
3324 }
3325 }
3326 return NULL;
3327 }
3328
3329 //------------------------------remove_pack_at---------------------------
3330 // Remove the pack at position pos in the packset
3331 void SuperWord::remove_pack_at(int pos) {
3332 Node_List* p = _packset.at(pos);
3333 for (uint i = 0; i < p->size(); i++) {
3334 Node* s = p->at(i);
3335 set_my_pack(s, NULL);
3336 }
3337 _packset.remove_at(pos);
3338 }
3339
3340 void SuperWord::packset_sort(int n) {
3341 // simple bubble sort so that we capitalize with O(n) when its already sorted
3342 while (n != 0) {
3343 bool swapped = false;
3344 for (int i = 1; i < n; i++) {
3345 Node_List* q_low = _packset.at(i-1);
3346 Node_List* q_i = _packset.at(i);
3347
3348 // only swap when we find something to swap
3349 if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
3350 Node_List* t = q_i;
3351 *(_packset.adr_at(i)) = q_low;
3352 *(_packset.adr_at(i-1)) = q_i;
3353 swapped = true;
3354 }
3355 }
3356 if (swapped == false) break;
3357 n--;
3358 }
3359 }
3360
3361 //------------------------------executed_first---------------------------
3362 // Return the node executed first in pack p. Uses the RPO block list
3363 // to determine order.
3364 Node* SuperWord::executed_first(Node_List* p) {
3365 Node* n = p->at(0);
3366 int n_rpo = bb_idx(n);
3367 for (uint i = 1; i < p->size(); i++) {
3368 Node* s = p->at(i);
3369 int s_rpo = bb_idx(s);
3370 if (s_rpo < n_rpo) {
3371 n = s;
3372 n_rpo = s_rpo;
3373 }
3374 }
3375 return n;
3376 }
3377
3378 //------------------------------executed_last---------------------------
3379 // Return the node executed last in pack p.
3380 Node* SuperWord::executed_last(Node_List* p) {
3381 Node* n = p->at(0);
3382 int n_rpo = bb_idx(n);
3383 for (uint i = 1; i < p->size(); i++) {
3384 Node* s = p->at(i);
3385 int s_rpo = bb_idx(s);
3386 if (s_rpo > n_rpo) {
3387 n = s;
3388 n_rpo = s_rpo;
3389 }
3390 }
3391 return n;
3392 }
3393
3394 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
3395 LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
3396 for (uint i = 0; i < p->size(); i++) {
3397 Node* n = p->at(i);
3398 assert(n->is_Load(), "only meaningful for loads");
3399 if (!n->depends_only_on_test()) {
3400 if (n->as_Load()->has_unknown_control_dependency() &&
3401 dep != LoadNode::Pinned) {
3402 // Upgrade to unknown control...
3403 dep = LoadNode::UnknownControl;
3404 } else {
3405 // Otherwise, we must pin it.
3406 dep = LoadNode::Pinned;
3407 }
3408 }
3409 }
3410 return dep;
3411 }
3412
3413
3414 //----------------------------align_initial_loop_index---------------------------
3415 // Adjust pre-loop limit so that in main loop, a load/store reference
3416 // to align_to_ref will be a position zero in the vector.
3417 // (iv + k) mod vector_align == 0
3418 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
3419 CountedLoopNode *main_head = lp()->as_CountedLoop();
3420 assert(main_head->is_main_loop(), "");
3421 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
3422 assert(pre_end != NULL, "we must have a correct pre-loop");
3423 Node *pre_opaq1 = pre_end->limit();
3424 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
3425 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
3426 Node *lim0 = pre_opaq->in(1);
3427
3428 // Where we put new limit calculations
3429 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
3430
3431 // Ensure the original loop limit is available from the
3432 // pre-loop Opaque1 node.
3433 Node *orig_limit = pre_opaq->original_loop_limit();
3434 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
3435
3436 SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
3437 assert(align_to_ref_p.valid(), "sanity");
3438
3439 // Given:
3440 // lim0 == original pre loop limit
3441 // V == v_align (power of 2)
3442 // invar == extra invariant piece of the address expression
3443 // e == offset [ +/- invar ]
3444 //
3445 // When reassociating expressions involving '%' the basic rules are:
3446 // (a - b) % k == 0 => a % k == b % k
3447 // and:
3448 // (a + b) % k == 0 => a % k == (k - b) % k
3449 //
3450 // For stride > 0 && scale > 0,
3451 // Derive the new pre-loop limit "lim" such that the two constraints:
3452 // (1) lim = lim0 + N (where N is some positive integer < V)
3453 // (2) (e + lim) % V == 0
3454 // are true.
3455 //
3456 // Substituting (1) into (2),
3457 // (e + lim0 + N) % V == 0
3458 // solve for N:
3459 // N = (V - (e + lim0)) % V
3460 // substitute back into (1), so that new limit
3461 // lim = lim0 + (V - (e + lim0)) % V
3462 //
3463 // For stride > 0 && scale < 0
3464 // Constraints:
3465 // lim = lim0 + N
3466 // (e - lim) % V == 0
3467 // Solving for lim:
3468 // (e - lim0 - N) % V == 0
3469 // N = (e - lim0) % V
3470 // lim = lim0 + (e - lim0) % V
3471 //
3472 // For stride < 0 && scale > 0
3473 // Constraints:
3474 // lim = lim0 - N
3475 // (e + lim) % V == 0
3476 // Solving for lim:
3477 // (e + lim0 - N) % V == 0
3478 // N = (e + lim0) % V
3479 // lim = lim0 - (e + lim0) % V
3480 //
3481 // For stride < 0 && scale < 0
3482 // Constraints:
3483 // lim = lim0 - N
3484 // (e - lim) % V == 0
3485 // Solving for lim:
3486 // (e - lim0 + N) % V == 0
3487 // N = (V - (e - lim0)) % V
3488 // lim = lim0 - (V - (e - lim0)) % V
3489
3490 int vw = vector_width_in_bytes(align_to_ref);
3491 int stride = iv_stride();
3492 int scale = align_to_ref_p.scale_in_bytes();
3493 int elt_size = align_to_ref_p.memory_size();
3494 int v_align = vw / elt_size;
3495 assert(v_align > 1, "sanity");
3496 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
3497 Node *offsn = _igvn.intcon(offset);
3498
3499 Node *e = offsn;
3500 if (align_to_ref_p.invar() != NULL) {
3501 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3502 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3503 Node* invar = align_to_ref_p.invar();
3504 if (_igvn.type(invar)->isa_long()) {
3505 // Computations are done % (vector width/element size) so it's
3506 // safe to simply convert invar to an int and loose the upper 32
3507 // bit half.
3508 invar = new ConvL2INode(invar);
3509 _igvn.register_new_node_with_optimizer(invar);
3510 }
3511 Node* aref = new URShiftINode(invar, log2_elt);
3512 _igvn.register_new_node_with_optimizer(aref);
3513 _phase->set_ctrl(aref, pre_ctrl);
3514 if (align_to_ref_p.negate_invar()) {
3515 e = new SubINode(e, aref);
3516 } else {
3517 e = new AddINode(e, aref);
3518 }
3519 _igvn.register_new_node_with_optimizer(e);
3520 _phase->set_ctrl(e, pre_ctrl);
3521 }
3522 if (vw > ObjectAlignmentInBytes || align_to_ref_p.base()->is_top()) {
3523 // incorporate base e +/- base && Mask >>> log2(elt)
3524 Node* xbase = new CastP2XNode(NULL, align_to_ref_p.adr());
3525 _igvn.register_new_node_with_optimizer(xbase);
3526 #ifdef _LP64
3527 xbase = new ConvL2INode(xbase);
3528 _igvn.register_new_node_with_optimizer(xbase);
3529 #endif
3530 Node* mask = _igvn.intcon(vw-1);
3531 Node* masked_xbase = new AndINode(xbase, mask);
3532 _igvn.register_new_node_with_optimizer(masked_xbase);
3533 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3534 Node* bref = new URShiftINode(masked_xbase, log2_elt);
3535 _igvn.register_new_node_with_optimizer(bref);
3536 _phase->set_ctrl(bref, pre_ctrl);
3537 e = new AddINode(e, bref);
3538 _igvn.register_new_node_with_optimizer(e);
3539 _phase->set_ctrl(e, pre_ctrl);
3540 }
3541
3542 // compute e +/- lim0
3543 if (scale < 0) {
3544 e = new SubINode(e, lim0);
3545 } else {
3546 e = new AddINode(e, lim0);
3547 }
3548 _igvn.register_new_node_with_optimizer(e);
3549 _phase->set_ctrl(e, pre_ctrl);
3550
3551 if (stride * scale > 0) {
3552 // compute V - (e +/- lim0)
3553 Node* va = _igvn.intcon(v_align);
3554 e = new SubINode(va, e);
3555 _igvn.register_new_node_with_optimizer(e);
3556 _phase->set_ctrl(e, pre_ctrl);
3557 }
3558 // compute N = (exp) % V
3559 Node* va_msk = _igvn.intcon(v_align - 1);
3560 Node* N = new AndINode(e, va_msk);
3561 _igvn.register_new_node_with_optimizer(N);
3562 _phase->set_ctrl(N, pre_ctrl);
3563
3564 // substitute back into (1), so that new limit
3565 // lim = lim0 + N
3566 Node* lim;
3567 if (stride < 0) {
3568 lim = new SubINode(lim0, N);
3569 } else {
3570 lim = new AddINode(lim0, N);
3571 }
3572 _igvn.register_new_node_with_optimizer(lim);
3573 _phase->set_ctrl(lim, pre_ctrl);
3574 Node* constrained =
3575 (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3576 : (Node*) new MaxINode(lim, orig_limit);
3577 _igvn.register_new_node_with_optimizer(constrained);
3578 _phase->set_ctrl(constrained, pre_ctrl);
3579 _igvn.replace_input_of(pre_opaq, 1, constrained);
3580 }
3581
3582 //----------------------------get_pre_loop_end---------------------------
3583 // Find pre loop end from main loop. Returns null if none.
3584 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3585 // The loop cannot be optimized if the graph shape at
3586 // the loop entry is inappropriate.
3587 if (!PhaseIdealLoop::is_canonical_loop_entry(cl)) {
3588 return NULL;
3589 }
3590
3591 Node* p_f = cl->skip_predicates()->in(0)->in(0);
3592 if (!p_f->is_IfFalse()) return NULL;
3593 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3594 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3595 CountedLoopNode* loop_node = pre_end->loopnode();
3596 if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3597 return pre_end;
3598 }
3599
3600 //------------------------------init---------------------------
3601 void SuperWord::init() {
3602 _dg.init();
3603 _packset.clear();
3604 _disjoint_ptrs.clear();
3605 _block.clear();
3606 _post_block.clear();
3607 _data_entry.clear();
3608 _mem_slice_head.clear();
3609 _mem_slice_tail.clear();
3610 _iteration_first.clear();
3611 _iteration_last.clear();
3612 _node_info.clear();
3613 _align_to_ref = NULL;
3614 _lpt = NULL;
3615 _lp = NULL;
3616 _bb = NULL;
3617 _iv = NULL;
3618 _race_possible = 0;
3619 _early_return = false;
3620 _num_work_vecs = 0;
3621 _num_reductions = 0;
3622 }
3623
3624 //------------------------------restart---------------------------
3625 void SuperWord::restart() {
3626 _dg.init();
3627 _packset.clear();
3628 _disjoint_ptrs.clear();
3629 _block.clear();
3630 _post_block.clear();
3631 _data_entry.clear();
3632 _mem_slice_head.clear();
3633 _mem_slice_tail.clear();
3634 _node_info.clear();
3635 }
3636
3637 //------------------------------print_packset---------------------------
3638 void SuperWord::print_packset() {
3639 #ifndef PRODUCT
3640 tty->print_cr("packset");
3641 for (int i = 0; i < _packset.length(); i++) {
3642 tty->print_cr("Pack: %d", i);
3643 Node_List* p = _packset.at(i);
3644 print_pack(p);
3645 }
3646 #endif
3647 }
3648
3649 //------------------------------print_pack---------------------------
3650 void SuperWord::print_pack(Node_List* p) {
3651 for (uint i = 0; i < p->size(); i++) {
3652 print_stmt(p->at(i));
3653 }
3654 }
3655
3656 //------------------------------print_bb---------------------------
3657 void SuperWord::print_bb() {
3658 #ifndef PRODUCT
3659 tty->print_cr("\nBlock");
3660 for (int i = 0; i < _block.length(); i++) {
3661 Node* n = _block.at(i);
3662 tty->print("%d ", i);
3663 if (n) {
3664 n->dump();
3665 }
3666 }
3667 #endif
3668 }
3669
3670 //------------------------------print_stmt---------------------------
3671 void SuperWord::print_stmt(Node* s) {
3672 #ifndef PRODUCT
3673 tty->print(" align: %d \t", alignment(s));
3674 s->dump();
3675 #endif
3676 }
3677
3678 //------------------------------blank---------------------------
3679 char* SuperWord::blank(uint depth) {
3680 static char blanks[101];
3681 assert(depth < 101, "too deep");
3682 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3683 blanks[depth] = '\0';
3684 return blanks;
3685 }
3686
3687
3688 //==============================SWPointer===========================
3689 #ifndef PRODUCT
3690 int SWPointer::Tracer::_depth = 0;
3691 #endif
3692 //----------------------------SWPointer------------------------
3693 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3694 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
3695 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3696 _nstack(nstack), _analyze_only(analyze_only),
3697 _stack_idx(0)
3698 #ifndef PRODUCT
3699 , _tracer(slp)
3700 #endif
3701 {
3702 NOT_PRODUCT(_tracer.ctor_1(mem);)
3703
3704 Node* adr = mem->in(MemNode::Address);
3705 if (!adr->is_AddP()) {
3706 assert(!valid(), "too complex");
3707 return;
3708 }
3709 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3710 Node* base = adr->in(AddPNode::Base);
3711 // The base address should be loop invariant
3712 if (!invariant(base)) {
3713 assert(!valid(), "base address is loop variant");
3714 return;
3715 }
3716 // unsafe references require misaligned vector access support
3717 if (base->is_top() && !Matcher::misaligned_vectors_ok()) {
3718 assert(!valid(), "unsafe access");
3719 return;
3720 }
3721
3722 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3723 NOT_PRODUCT(_tracer.ctor_2(adr);)
3724
3725 int i;
3726 for (i = 0; i < 3; i++) {
3727 NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3728
3729 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3730 assert(!valid(), "too complex");
3731 return;
3732 }
3733 adr = adr->in(AddPNode::Address);
3734 NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3735
3736 if (base == adr || !adr->is_AddP()) {
3737 NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3738 break; // stop looking at addp's
3739 }
3740 }
3741 if (!invariant(adr)) {
3742 assert(!valid(), "adr is loop variant");
3743 return;
3744 }
3745
3746 if (!base->is_top() && adr != base) {
3747 assert(!valid(), "adr and base differ");
3748 return;
3749 }
3750
3751 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3752 NOT_PRODUCT(_tracer.ctor_6(mem);)
3753
3754 _base = base;
3755 _adr = adr;
3756 assert(valid(), "Usable");
3757 }
3758
3759 // Following is used to create a temporary object during
3760 // the pattern match of an address expression.
3761 SWPointer::SWPointer(SWPointer* p) :
3762 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
3763 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3764 _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3765 _stack_idx(p->_stack_idx)
3766 #ifndef PRODUCT
3767 , _tracer(p->_slp)
3768 #endif
3769 {}
3770
3771
3772 bool SWPointer::invariant(Node* n) {
3773 NOT_PRODUCT(Tracer::Depth dd;)
3774 Node *n_c = phase()->get_ctrl(n);
3775 NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3776 return !lpt()->is_member(phase()->get_loop(n_c));
3777 }
3778
3779 //------------------------scaled_iv_plus_offset--------------------
3780 // Match: k*iv + offset
3781 // where: k is a constant that maybe zero, and
3782 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
3783 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3784 NOT_PRODUCT(Tracer::Depth ddd;)
3785 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3786
3787 if (scaled_iv(n)) {
3788 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3789 return true;
3790 }
3791
3792 if (offset_plus_k(n)) {
3793 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3794 return true;
3795 }
3796
3797 int opc = n->Opcode();
3798 if (opc == Op_AddI) {
3799 if (offset_plus_k(n->in(2)) && scaled_iv_plus_offset(n->in(1))) {
3800 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3801 return true;
3802 }
3803 if (offset_plus_k(n->in(1)) && scaled_iv_plus_offset(n->in(2))) {
3804 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3805 return true;
3806 }
3807 } else if (opc == Op_SubI) {
3808 if (offset_plus_k(n->in(2), true) && scaled_iv_plus_offset(n->in(1))) {
3809 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3810 return true;
3811 }
3812 if (offset_plus_k(n->in(1)) && scaled_iv_plus_offset(n->in(2))) {
3813 _scale *= -1;
3814 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3815 return true;
3816 }
3817 }
3818
3819 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3820 return false;
3821 }
3822
3823 //----------------------------scaled_iv------------------------
3824 // Match: k*iv where k is a constant that's not zero
3825 bool SWPointer::scaled_iv(Node* n) {
3826 NOT_PRODUCT(Tracer::Depth ddd;)
3827 NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3828
3829 if (_scale != 0) { // already found a scale
3830 NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3831 return false;
3832 }
3833
3834 if (n == iv()) {
3835 _scale = 1;
3836 NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3837 return true;
3838 }
3839 if (_analyze_only && (invariant(n) == false)) {
3840 _nstack->push(n, _stack_idx++);
3841 }
3842
3843 int opc = n->Opcode();
3844 if (opc == Op_MulI) {
3845 if (n->in(1) == iv() && n->in(2)->is_Con()) {
3846 _scale = n->in(2)->get_int();
3847 NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3848 return true;
3849 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3850 _scale = n->in(1)->get_int();
3851 NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3852 return true;
3853 }
3854 } else if (opc == Op_LShiftI) {
3855 if (n->in(1) == iv() && n->in(2)->is_Con()) {
3856 _scale = 1 << n->in(2)->get_int();
3857 NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3858 return true;
3859 }
3860 } else if (opc == Op_ConvI2L) {
3861 if (n->in(1)->Opcode() == Op_CastII &&
3862 n->in(1)->as_CastII()->has_range_check()) {
3863 // Skip range check dependent CastII nodes
3864 n = n->in(1);
3865 }
3866 if (scaled_iv_plus_offset(n->in(1))) {
3867 NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3868 return true;
3869 }
3870 } else if (opc == Op_LShiftL) {
3871 if (!has_iv() && _invar == NULL) {
3872 // Need to preserve the current _offset value, so
3873 // create a temporary object for this expression subtree.
3874 // Hacky, so should re-engineer the address pattern match.
3875 NOT_PRODUCT(Tracer::Depth dddd;)
3876 SWPointer tmp(this);
3877 NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3878
3879 if (tmp.scaled_iv_plus_offset(n->in(1))) {
3880 if (tmp._invar == NULL || _slp->do_vector_loop()) {
3881 int mult = 1 << n->in(2)->get_int();
3882 _scale = tmp._scale * mult;
3883 _offset += tmp._offset * mult;
3884 _invar = tmp._invar;
3885 NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3886 return true;
3887 }
3888 }
3889 }
3890 }
3891 NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3892 return false;
3893 }
3894
3895 //----------------------------offset_plus_k------------------------
3896 // Match: offset is (k [+/- invariant])
3897 // where k maybe zero and invariant is optional, but not both.
3898 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3899 NOT_PRODUCT(Tracer::Depth ddd;)
3900 NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3901
3902 int opc = n->Opcode();
3903 if (opc == Op_ConI) {
3904 _offset += negate ? -(n->get_int()) : n->get_int();
3905 NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3906 return true;
3907 } else if (opc == Op_ConL) {
3908 // Okay if value fits into an int
3909 const TypeLong* t = n->find_long_type();
3910 if (t->higher_equal(TypeLong::INT)) {
3911 jlong loff = n->get_long();
3912 jint off = (jint)loff;
3913 _offset += negate ? -off : loff;
3914 NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3915 return true;
3916 }
3917 NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3918 return false;
3919 }
3920 if (_invar != NULL) { // already has an invariant
3921 NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3922 return false;
3923 }
3924
3925 if (_analyze_only && (invariant(n) == false)) {
3926 _nstack->push(n, _stack_idx++);
3927 }
3928 if (opc == Op_AddI) {
3929 if (n->in(2)->is_Con() && invariant(n->in(1))) {
3930 _negate_invar = negate;
3931 _invar = n->in(1);
3932 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3933 NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3934 return true;
3935 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3936 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3937 _negate_invar = negate;
3938 _invar = n->in(2);
3939 NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3940 return true;
3941 }
3942 }
3943 if (opc == Op_SubI) {
3944 if (n->in(2)->is_Con() && invariant(n->in(1))) {
3945 _negate_invar = negate;
3946 _invar = n->in(1);
3947 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3948 NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3949 return true;
3950 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3951 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3952 _negate_invar = !negate;
3953 _invar = n->in(2);
3954 NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3955 return true;
3956 }
3957 }
3958 if (invariant(n)) {
3959 if (opc == Op_ConvI2L) {
3960 n = n->in(1);
3961 if (n->Opcode() == Op_CastII &&
3962 n->as_CastII()->has_range_check()) {
3963 // Skip range check dependent CastII nodes
3964 assert(invariant(n), "sanity");
3965 n = n->in(1);
3966 }
3967 }
3968 _negate_invar = negate;
3969 _invar = n;
3970 NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3971 return true;
3972 }
3973
3974 NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3975 return false;
3976 }
3977
3978 //----------------------------print------------------------
3979 void SWPointer::print() {
3980 #ifndef PRODUCT
3981 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
3982 _base != NULL ? _base->_idx : 0,
3983 _adr != NULL ? _adr->_idx : 0,
3984 _scale, _offset,
3985 _negate_invar?'-':'+',
3986 _invar != NULL ? _invar->_idx : 0);
3987 #endif
3988 }
3989
3990 //----------------------------tracing------------------------
3991 #ifndef PRODUCT
3992 void SWPointer::Tracer::print_depth() {
3993 for (int ii = 0; ii<_depth; ++ii) tty->print(" ");
3994 }
3995
3996 void SWPointer::Tracer::ctor_1 (Node* mem) {
3997 if(_slp->is_trace_alignment()) {
3998 print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3999 }
4000 }
4001
4002 void SWPointer::Tracer::ctor_2(Node* adr) {
4003 if(_slp->is_trace_alignment()) {
4004 //store_depth();
4005 inc_depth();
4006 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
4007 inc_depth();
4008 print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
4009 }
4010 }
4011
4012 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
4013 if(_slp->is_trace_alignment()) {
4014 inc_depth();
4015 Node* offset = adr->in(AddPNode::Offset);
4016 print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
4017 }
4018 }
4019
4020 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
4021 if(_slp->is_trace_alignment()) {
4022 inc_depth();
4023 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
4024 }
4025 }
4026
4027 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
4028 if(_slp->is_trace_alignment()) {
4029 inc_depth();
4030 if (base == adr) {
4031 print_depth(); tty->print_cr(" \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
4032 } else if (!adr->is_AddP()) {
4033 print_depth(); tty->print_cr(" \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
4034 }
4035 }
4036 }
4037
4038 void SWPointer::Tracer::ctor_6(Node* mem) {
4039 if(_slp->is_trace_alignment()) {
4040 //restore_depth();
4041 print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
4042 }
4043 }
4044
4045 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
4046 if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
4047 int is_member = _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
4048 int in_bb = _slp->in_bb(n);
4049 print_depth(); tty->print(" \\ "); tty->print_cr(" %d SWPointer::invariant conditions differ: n_c %d", n->_idx, n_c->_idx);
4050 print_depth(); tty->print(" \\ "); tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
4051 print_depth(); tty->print(" \\ "); n->dump();
4052 print_depth(); tty->print(" \\ "); n_c->dump();
4053 }
4054 }
4055
4056 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
4057 if(_slp->is_trace_alignment()) {
4058 print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
4059 n->dump();
4060 }
4061 }
4062
4063 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
4064 if(_slp->is_trace_alignment()) {
4065 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
4066 }
4067 }
4068
4069 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
4070 if(_slp->is_trace_alignment()) {
4071 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
4072 }
4073 }
4074
4075 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
4076 if(_slp->is_trace_alignment()) {
4077 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
4078 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
4079 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
4080 }
4081 }
4082
4083 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
4084 if(_slp->is_trace_alignment()) {
4085 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
4086 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
4087 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
4088 }
4089 }
4090
4091 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
4092 if(_slp->is_trace_alignment()) {
4093 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
4094 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
4095 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
4096 }
4097 }
4098
4099 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
4100 if(_slp->is_trace_alignment()) {
4101 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
4102 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
4103 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
4104 }
4105 }
4106
4107 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
4108 if(_slp->is_trace_alignment()) {
4109 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
4110 }
4111 }
4112
4113 void SWPointer::Tracer::scaled_iv_1(Node* n) {
4114 if(_slp->is_trace_alignment()) {
4115 print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
4116 }
4117 }
4118
4119 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
4120 if(_slp->is_trace_alignment()) {
4121 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
4122 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
4123 }
4124 }
4125
4126 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
4127 if(_slp->is_trace_alignment()) {
4128 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
4129 }
4130 }
4131
4132 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
4133 if(_slp->is_trace_alignment()) {
4134 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
4135 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
4136 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4137 }
4138 }
4139
4140 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
4141 if(_slp->is_trace_alignment()) {
4142 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
4143 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
4144 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4145 }
4146 }
4147
4148 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
4149 if(_slp->is_trace_alignment()) {
4150 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
4151 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
4152 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4153 }
4154 }
4155
4156 void SWPointer::Tracer::scaled_iv_7(Node* n) {
4157 if(_slp->is_trace_alignment()) {
4158 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
4159 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
4160 inc_depth(); inc_depth();
4161 print_depth(); n->in(1)->dump();
4162 dec_depth(); dec_depth();
4163 }
4164 }
4165
4166 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
4167 if(_slp->is_trace_alignment()) {
4168 print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
4169 }
4170 }
4171
4172 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
4173 if(_slp->is_trace_alignment()) {
4174 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
4175 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",
4176 n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
4177 inc_depth(); inc_depth();
4178 print_depth(); n->in(1)->dump();
4179 print_depth(); n->in(2)->dump();
4180 dec_depth(); dec_depth();
4181 }
4182 }
4183
4184 void SWPointer::Tracer::scaled_iv_10(Node* n) {
4185 if(_slp->is_trace_alignment()) {
4186 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
4187 }
4188 }
4189
4190 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
4191 if(_slp->is_trace_alignment()) {
4192 print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
4193 }
4194 }
4195
4196 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
4197 if(_slp->is_trace_alignment()) {
4198 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
4199 }
4200 }
4201
4202 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
4203 if(_slp->is_trace_alignment()) {
4204 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
4205 }
4206 }
4207
4208 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
4209 if(_slp->is_trace_alignment()) {
4210 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4211 print_depth(); tty->print_cr(" \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
4212 }
4213 }
4214
4215 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
4216 if(_slp->is_trace_alignment()) {
4217 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
4218 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
4219 }
4220 }
4221
4222 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4223 if(_slp->is_trace_alignment()) {
4224 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4225 n->_idx, _negate_invar, _invar->_idx, _offset);
4226 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4227 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4228 }
4229 }
4230
4231 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4232 if(_slp->is_trace_alignment()) {
4233 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4234 n->_idx, _negate_invar, _invar->_idx, _offset);
4235 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4236 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4237 }
4238 }
4239
4240 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4241 if(_slp->is_trace_alignment()) {
4242 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4243 n->_idx, _negate_invar, _invar->_idx, _offset);
4244 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4245 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4246 }
4247 }
4248
4249 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4250 if(_slp->is_trace_alignment()) {
4251 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);
4252 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4253 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4254 }
4255 }
4256
4257 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4258 if(_slp->is_trace_alignment()) {
4259 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);
4260 print_depth(); tty->print_cr(" \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
4261 }
4262 }
4263
4264 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
4265 if(_slp->is_trace_alignment()) {
4266 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4267 }
4268 }
4269
4270 #endif
4271 // ========================= OrderedPair =====================
4272
4273 const OrderedPair OrderedPair::initial;
4274
4275 // ========================= SWNodeInfo =====================
4276
4277 const SWNodeInfo SWNodeInfo::initial;
4278
4279
4280 // ============================ DepGraph ===========================
4281
4282 //------------------------------make_node---------------------------
4283 // Make a new dependence graph node for an ideal node.
4284 DepMem* DepGraph::make_node(Node* node) {
4285 DepMem* m = new (_arena) DepMem(node);
4286 if (node != NULL) {
4287 assert(_map.at_grow(node->_idx) == NULL, "one init only");
4288 _map.at_put_grow(node->_idx, m);
4289 }
4290 return m;
4291 }
4292
4293 //------------------------------make_edge---------------------------
4294 // Make a new dependence graph edge from dpred -> dsucc
4295 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
4296 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
4297 dpred->set_out_head(e);
4298 dsucc->set_in_head(e);
4299 return e;
4300 }
4301
4302 // ========================== DepMem ========================
4303
4304 //------------------------------in_cnt---------------------------
4305 int DepMem::in_cnt() {
4306 int ct = 0;
4307 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
4308 return ct;
4309 }
4310
4311 //------------------------------out_cnt---------------------------
4312 int DepMem::out_cnt() {
4313 int ct = 0;
4314 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
4315 return ct;
4316 }
4317
4318 //------------------------------print-----------------------------
4319 void DepMem::print() {
4320 #ifndef PRODUCT
4321 tty->print(" DepNode %d (", _node->_idx);
4322 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
4323 Node* pred = p->pred()->node();
4324 tty->print(" %d", pred != NULL ? pred->_idx : 0);
4325 }
4326 tty->print(") [");
4327 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
4328 Node* succ = s->succ()->node();
4329 tty->print(" %d", succ != NULL ? succ->_idx : 0);
4330 }
4331 tty->print_cr(" ]");
4332 #endif
4333 }
4334
4335 // =========================== DepEdge =========================
4336
4337 //------------------------------DepPreds---------------------------
4338 void DepEdge::print() {
4339 #ifndef PRODUCT
4340 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
4341 #endif
4342 }
4343
4344 // =========================== DepPreds =========================
4345 // Iterator over predecessor edges in the dependence graph.
4346
4347 //------------------------------DepPreds---------------------------
4348 DepPreds::DepPreds(Node* n, DepGraph& dg) {
4349 _n = n;
4350 _done = false;
4351 if (_n->is_Store() || _n->is_Load()) {
4352 _next_idx = MemNode::Address;
4353 _end_idx = n->req();
4354 _dep_next = dg.dep(_n)->in_head();
4355 } else if (_n->is_Mem()) {
4356 _next_idx = 0;
4357 _end_idx = 0;
4358 _dep_next = dg.dep(_n)->in_head();
4359 } else {
4360 _next_idx = 1;
4361 _end_idx = _n->req();
4362 _dep_next = NULL;
4363 }
4364 next();
4365 }
4366
4367 //------------------------------next---------------------------
4368 void DepPreds::next() {
4369 if (_dep_next != NULL) {
4370 _current = _dep_next->pred()->node();
4371 _dep_next = _dep_next->next_in();
4372 } else if (_next_idx < _end_idx) {
4373 _current = _n->in(_next_idx++);
4374 } else {
4375 _done = true;
4376 }
4377 }
4378
4379 // =========================== DepSuccs =========================
4380 // Iterator over successor edges in the dependence graph.
4381
4382 //------------------------------DepSuccs---------------------------
4383 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
4384 _n = n;
4385 _done = false;
4386 if (_n->is_Load()) {
4387 _next_idx = 0;
4388 _end_idx = _n->outcnt();
4389 _dep_next = dg.dep(_n)->out_head();
4390 } else if (_n->is_Mem() || (_n->is_Phi() && _n->bottom_type() == Type::MEMORY)) {
4391 _next_idx = 0;
4392 _end_idx = 0;
4393 _dep_next = dg.dep(_n)->out_head();
4394 } else {
4395 _next_idx = 0;
4396 _end_idx = _n->outcnt();
4397 _dep_next = NULL;
4398 }
4399 next();
4400 }
4401
4402 //-------------------------------next---------------------------
4403 void DepSuccs::next() {
4404 if (_dep_next != NULL) {
4405 _current = _dep_next->succ()->node();
4406 _dep_next = _dep_next->next_out();
4407 } else if (_next_idx < _end_idx) {
4408 _current = _n->raw_out(_next_idx++);
4409 } else {
4410 _done = true;
4411 }
4412 }
4413
4414 //
4415 // --------------------------------- vectorization/simd -----------------------------------
4416 //
4417 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
4418 return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
4419 }
4420 bool SuperWord::same_generation(Node* a, Node* b) const {
4421 return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
4422 }
4423
4424 Node* SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
4425 assert(in_bb(ld), "must be in block");
4426 if (_clone_map.gen(ld->_idx) == _ii_first) {
4427 #ifndef PRODUCT
4428 if (_vector_loop_debug) {
4429 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
4430 _clone_map.gen(ld->_idx));
4431 }
4432 #endif
4433 return NULL; //we think that any ld in the first gen being vectorizable
4434 }
4435
4436 Node* mem = ld->in(MemNode::Memory);
4437 if (mem->outcnt() <= 1) {
4438 // we don't want to remove the only edge from mem node to load
4439 #ifndef PRODUCT
4440 if (_vector_loop_debug) {
4441 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",
4442 mem->_idx, ld->_idx);
4443 ld->dump();
4444 mem->dump();
4445 }
4446 #endif
4447 return NULL;
4448 }
4449 if (!in_bb(mem) || same_generation(mem, ld)) {
4450 #ifndef PRODUCT
4451 if (_vector_loop_debug) {
4452 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
4453 _clone_map.gen(mem->_idx));
4454 }
4455 #endif
4456 return NULL; // does not depend on loop volatile node or depends on the same generation
4457 }
4458
4459 //otherwise first node should depend on mem-phi
4460 Node* first = first_node(ld);
4461 assert(first->is_Load(), "must be Load");
4462 Node* phi = first->as_Load()->in(MemNode::Memory);
4463 if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
4464 #ifndef PRODUCT
4465 if (_vector_loop_debug) {
4466 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");
4467 ld->dump();
4468 first->dump();
4469 }
4470 #endif
4471 return NULL;
4472 }
4473
4474 Node* tail = 0;
4475 for (int m = 0; m < _mem_slice_head.length(); m++) {
4476 if (_mem_slice_head.at(m) == phi) {
4477 tail = _mem_slice_tail.at(m);
4478 }
4479 }
4480 if (tail == 0) { //test that found phi is in the list _mem_slice_head
4481 #ifndef PRODUCT
4482 if (_vector_loop_debug) {
4483 tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
4484 ld->_idx, phi->_idx);
4485 ld->dump();
4486 phi->dump();
4487 }
4488 #endif
4489 return NULL;
4490 }
4491
4492 // now all conditions are met
4493 return phi;
4494 }
4495
4496 Node* SuperWord::first_node(Node* nd) {
4497 for (int ii = 0; ii < _iteration_first.length(); ii++) {
4498 Node* nnn = _iteration_first.at(ii);
4499 if (same_origin_idx(nnn, nd)) {
4500 #ifndef PRODUCT
4501 if (_vector_loop_debug) {
4502 tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
4503 nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
4504 }
4505 #endif
4506 return nnn;
4507 }
4508 }
4509
4510 #ifndef PRODUCT
4511 if (_vector_loop_debug) {
4512 tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4513 nd->_idx, _clone_map.idx(nd->_idx));
4514 }
4515 #endif
4516 return 0;
4517 }
4518
4519 Node* SuperWord::last_node(Node* nd) {
4520 for (int ii = 0; ii < _iteration_last.length(); ii++) {
4521 Node* nnn = _iteration_last.at(ii);
4522 if (same_origin_idx(nnn, nd)) {
4523 #ifndef PRODUCT
4524 if (_vector_loop_debug) {
4525 tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4526 _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4527 }
4528 #endif
4529 return nnn;
4530 }
4531 }
4532 return 0;
4533 }
4534
4535 int SuperWord::mark_generations() {
4536 Node *ii_err = NULL, *tail_err = NULL;
4537 for (int i = 0; i < _mem_slice_head.length(); i++) {
4538 Node* phi = _mem_slice_head.at(i);
4539 assert(phi->is_Phi(), "must be phi");
4540
4541 Node* tail = _mem_slice_tail.at(i);
4542 if (_ii_last == -1) {
4543 tail_err = tail;
4544 _ii_last = _clone_map.gen(tail->_idx);
4545 }
4546 else if (_ii_last != _clone_map.gen(tail->_idx)) {
4547 #ifndef PRODUCT
4548 if (TraceSuperWord && Verbose) {
4549 tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4550 tail->dump();
4551 tail_err->dump();
4552 }
4553 #endif
4554 return -1;
4555 }
4556
4557 // find first iteration in the loop
4558 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4559 Node* ii = phi->fast_out(i);
4560 if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4561 if (_ii_first == -1) {
4562 ii_err = ii;
4563 _ii_first = _clone_map.gen(ii->_idx);
4564 } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4565 #ifndef PRODUCT
4566 if (TraceSuperWord && Verbose) {
4567 tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4568 ii->dump();
4569 if (ii_err!= 0) {
4570 ii_err->dump();
4571 }
4572 }
4573 #endif
4574 return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4575 }
4576 }
4577 }//for (DUIterator_Fast imax,
4578 }//for (int i...
4579
4580 if (_ii_first == -1 || _ii_last == -1) {
4581 if (TraceSuperWord && Verbose) {
4582 tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4583 }
4584 return -1; // something vent wrong
4585 }
4586 // collect nodes in the first and last generations
4587 assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4588 assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4589 for (int j = 0; j < _block.length(); j++) {
4590 Node* n = _block.at(j);
4591 node_idx_t gen = _clone_map.gen(n->_idx);
4592 if ((signed)gen == _ii_first) {
4593 _iteration_first.push(n);
4594 } else if ((signed)gen == _ii_last) {
4595 _iteration_last.push(n);
4596 }
4597 }
4598
4599 // building order of iterations
4600 if (_ii_order.length() == 0 && ii_err != 0) {
4601 assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4602 Node* nd = ii_err;
4603 while(_clone_map.gen(nd->_idx) != _ii_last) {
4604 _ii_order.push(_clone_map.gen(nd->_idx));
4605 bool found = false;
4606 for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4607 Node* use = nd->fast_out(i);
4608 if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4609 found = true;
4610 nd = use;
4611 break;
4612 }
4613 }//for
4614
4615 if (found == false) {
4616 if (TraceSuperWord && Verbose) {
4617 tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4618 }
4619 _ii_order.clear();
4620 return -1;
4621 }
4622 } //while
4623 _ii_order.push(_clone_map.gen(nd->_idx));
4624 }
4625
4626 #ifndef PRODUCT
4627 if (_vector_loop_debug) {
4628 tty->print_cr("SuperWord::mark_generations");
4629 tty->print_cr("First generation (%d) nodes:", _ii_first);
4630 for (int ii = 0; ii < _iteration_first.length(); ii++) _iteration_first.at(ii)->dump();
4631 tty->print_cr("Last generation (%d) nodes:", _ii_last);
4632 for (int ii = 0; ii < _iteration_last.length(); ii++) _iteration_last.at(ii)->dump();
4633 tty->print_cr(" ");
4634
4635 tty->print("SuperWord::List of generations: ");
4636 for (int jj = 0; jj < _ii_order.length(); ++jj) {
4637 tty->print("%d:%d ", jj, _ii_order.at(jj));
4638 }
4639 tty->print_cr(" ");
4640 }
4641 #endif
4642
4643 return _ii_first;
4644 }
4645
4646 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4647 assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4648 assert(same_origin_idx(gold, fix), "should be clones of the same node");
4649 Node* gin1 = gold->in(1);
4650 Node* gin2 = gold->in(2);
4651 Node* fin1 = fix->in(1);
4652 Node* fin2 = fix->in(2);
4653 bool swapped = false;
4654
4655 if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4656 if (same_origin_idx(gin1, fin1) &&
4657 same_origin_idx(gin2, fin2)) {
4658 return true; // nothing to fix
4659 }
4660 if (same_origin_idx(gin1, fin2) &&
4661 same_origin_idx(gin2, fin1)) {
4662 fix->swap_edges(1, 2);
4663 swapped = true;
4664 }
4665 }
4666 // at least one input comes from outside of bb
4667 if (gin1->_idx == fin1->_idx) {
4668 return true; // nothing to fix
4669 }
4670 if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx)) { //swapping is expensive, check condition first
4671 fix->swap_edges(1, 2);
4672 swapped = true;
4673 }
4674
4675 if (swapped) {
4676 #ifndef PRODUCT
4677 if (_vector_loop_debug) {
4678 tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4679 }
4680 #endif
4681 return true;
4682 }
4683
4684 if (TraceSuperWord && Verbose) {
4685 tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4686 }
4687
4688 return false;
4689 }
4690
4691 bool SuperWord::pack_parallel() {
4692 #ifndef PRODUCT
4693 if (_vector_loop_debug) {
4694 tty->print_cr("SuperWord::pack_parallel: START");
4695 }
4696 #endif
4697
4698 _packset.clear();
4699
4700 for (int ii = 0; ii < _iteration_first.length(); ii++) {
4701 Node* nd = _iteration_first.at(ii);
4702 if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4703 Node_List* pk = new Node_List();
4704 pk->push(nd);
4705 for (int gen = 1; gen < _ii_order.length(); ++gen) {
4706 for (int kk = 0; kk < _block.length(); kk++) {
4707 Node* clone = _block.at(kk);
4708 if (same_origin_idx(clone, nd) &&
4709 _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4710 if (nd->is_Add() || nd->is_Mul()) {
4711 fix_commutative_inputs(nd, clone);
4712 }
4713 pk->push(clone);
4714 if (pk->size() == 4) {
4715 _packset.append(pk);
4716 #ifndef PRODUCT
4717 if (_vector_loop_debug) {
4718 tty->print_cr("SuperWord::pack_parallel: added pack ");
4719 pk->dump();
4720 }
4721 #endif
4722 if (_clone_map.gen(clone->_idx) != _ii_last) {
4723 pk = new Node_List();
4724 }
4725 }
4726 break;
4727 }
4728 }
4729 }//for
4730 }//if
4731 }//for
4732
4733 #ifndef PRODUCT
4734 if (_vector_loop_debug) {
4735 tty->print_cr("SuperWord::pack_parallel: END");
4736 }
4737 #endif
4738
4739 return true;
4740 }
4741
4742 bool SuperWord::hoist_loads_in_graph() {
4743 GrowableArray<Node*> loads;
4744
4745 #ifndef PRODUCT
4746 if (_vector_loop_debug) {
4747 tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4748 }
4749 #endif
4750
4751 for (int i = 0; i < _mem_slice_head.length(); i++) {
4752 Node* n = _mem_slice_head.at(i);
4753 if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4754 if (TraceSuperWord && Verbose) {
4755 tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4756 }
4757 continue;
4758 }
4759
4760 #ifndef PRODUCT
4761 if (_vector_loop_debug) {
4762 tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d = _mem_slice_head.at(%d);", n->_idx, i);
4763 }
4764 #endif
4765
4766 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4767 Node* ld = n->fast_out(i);
4768 if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4769 for (int i = 0; i < _block.length(); i++) {
4770 Node* ld2 = _block.at(i);
4771 if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4772 !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4773 #ifndef PRODUCT
4774 if (_vector_loop_debug) {
4775 tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4776 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4777 }
4778 #endif
4779 // could not do on-the-fly, since iterator is immutable
4780 loads.push(ld2);
4781 }
4782 }// for
4783 }//if
4784 }//for (DUIterator_Fast imax,
4785 }//for (int i = 0; i
4786
4787 for (int i = 0; i < loads.length(); i++) {
4788 LoadNode* ld = loads.at(i)->as_Load();
4789 Node* phi = find_phi_for_mem_dep(ld);
4790 if (phi != NULL) {
4791 #ifndef PRODUCT
4792 if (_vector_loop_debug) {
4793 tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4794 MemNode::Memory, ld->_idx, phi->_idx);
4795 }
4796 #endif
4797 _igvn.replace_input_of(ld, MemNode::Memory, phi);
4798 }
4799 }//for
4800
4801 restart(); // invalidate all basic structures, since we rebuilt the graph
4802
4803 if (TraceSuperWord && Verbose) {
4804 tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4805 }
4806
4807 return true;
4808 }