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