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