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