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