1 /*
2 * Copyright (c) 2007, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 */
23
24 #include "precompiled.hpp"
25 #include "compiler/compileLog.hpp"
26 #include "libadt/vectset.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/convertnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/matcher.hpp"
35 #include "opto/memnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/opcodes.hpp"
38 #include "opto/opaquenode.hpp"
39 #include "opto/superword.hpp"
40 #include "opto/vectornode.hpp"
41 #include "opto/movenode.hpp"
42 #include "utilities/powerOfTwo.hpp"
43
44 //
45 // S U P E R W O R D T R A N S F O R M
46 //=============================================================================
47
48 //------------------------------SuperWord---------------------------
SuperWord(PhaseIdealLoop * phase)49 SuperWord::SuperWord(PhaseIdealLoop* phase) :
50 _phase(phase),
51 _arena(phase->C->comp_arena()),
52 _igvn(phase->_igvn),
53 _packset(arena(), 8, 0, NULL), // packs for the current block
54 _bb_idx(arena(), (int)(1.10 * phase->C->unique()), 0, 0), // node idx to index in bb
55 _block(arena(), 8, 0, NULL), // nodes in current block
56 _post_block(arena(), 8, 0, NULL), // nodes common to current block which are marked as post loop vectorizable
57 _data_entry(arena(), 8, 0, NULL), // nodes with all inputs from outside
58 _mem_slice_head(arena(), 8, 0, NULL), // memory slice heads
59 _mem_slice_tail(arena(), 8, 0, NULL), // memory slice tails
60 _node_info(arena(), 8, 0, SWNodeInfo::initial), // info needed per node
61 _clone_map(phase->C->clone_map()), // map of nodes created in cloning
62 _cmovev_kit(_arena, this), // map to facilitate CMoveV creation
63 _align_to_ref(NULL), // memory reference to align vectors to
64 _disjoint_ptrs(arena(), 8, 0, OrderedPair::initial), // runtime disambiguated pointer pairs
65 _dg(_arena), // dependence graph
66 _visited(arena()), // visited node set
67 _post_visited(arena()), // post visited node set
68 _n_idx_list(arena(), 8), // scratch list of (node,index) pairs
69 _nlist(arena(), 8, 0, NULL), // scratch list of nodes
70 _stk(arena(), 8, 0, NULL), // scratch stack of nodes
71 _lpt(NULL), // loop tree node
72 _lp(NULL), // LoopNode
73 _bb(NULL), // basic block
74 _iv(NULL), // induction var
75 _race_possible(false), // cases where SDMU is true
76 _early_return(true), // analysis evaluations routine
77 _do_vector_loop(phase->C->do_vector_loop()), // whether to do vectorization/simd style
78 _do_reserve_copy(DoReserveCopyInSuperWord),
79 _num_work_vecs(0), // amount of vector work we have
80 _num_reductions(0), // amount of reduction work we have
81 _ii_first(-1), // first loop generation index - only if do_vector_loop()
82 _ii_last(-1), // last loop generation index - only if do_vector_loop()
83 _ii_order(arena(), 8, 0, 0)
84 {
85 #ifndef PRODUCT
86 _vector_loop_debug = 0;
87 if (_phase->C->method() != NULL) {
88 _vector_loop_debug = phase->C->directive()->VectorizeDebugOption;
89 }
90
91 #endif
92 }
93
94 //------------------------------transform_loop---------------------------
transform_loop(IdealLoopTree * lpt,bool do_optimization)95 void SuperWord::transform_loop(IdealLoopTree* lpt, bool do_optimization) {
96 assert(UseSuperWord, "should be");
97 // Do vectors exist on this architecture?
98 if (Matcher::vector_width_in_bytes(T_BYTE) < 2) return;
99
100 assert(lpt->_head->is_CountedLoop(), "must be");
101 CountedLoopNode *cl = lpt->_head->as_CountedLoop();
102
103 if (!cl->is_valid_counted_loop()) return; // skip malformed counted loop
104
105 bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
106 if (post_loop_allowed) {
107 if (cl->is_reduction_loop()) return; // no predication mapping
108 Node *limit = cl->limit();
109 if (limit->is_Con()) return; // non constant limits only
110 // Now check the limit for expressions we do not handle
111 if (limit->is_Add()) {
112 Node *in2 = limit->in(2);
113 if (in2->is_Con()) {
114 int val = in2->get_int();
115 // should not try to program these cases
116 if (val < 0) return;
117 }
118 }
119 }
120
121 // skip any loop that has not been assigned max unroll by analysis
122 if (do_optimization) {
123 if (SuperWordLoopUnrollAnalysis && cl->slp_max_unroll() == 0) return;
124 }
125
126 // Check for no control flow in body (other than exit)
127 Node *cl_exit = cl->loopexit();
128 if (cl->is_main_loop() && (cl_exit->in(0) != lpt->_head)) {
129 #ifndef PRODUCT
130 if (TraceSuperWord) {
131 tty->print_cr("SuperWord::transform_loop: loop too complicated, cl_exit->in(0) != lpt->_head");
132 tty->print("cl_exit %d", cl_exit->_idx); cl_exit->dump();
133 tty->print("cl_exit->in(0) %d", cl_exit->in(0)->_idx); cl_exit->in(0)->dump();
134 tty->print("lpt->_head %d", lpt->_head->_idx); lpt->_head->dump();
135 lpt->dump_head();
136 }
137 #endif
138 return;
139 }
140
141 // Make sure the are no extra control users of the loop backedge
142 if (cl->back_control()->outcnt() != 1) {
143 return;
144 }
145
146 // Skip any loops already optimized by slp
147 if (cl->is_vectorized_loop()) return;
148
149 if (cl->is_unroll_only()) return;
150
151 if (cl->is_main_loop()) {
152 // Check for pre-loop ending with CountedLoopEnd(Bool(Cmp(x,Opaque1(limit))))
153 CountedLoopEndNode* pre_end = get_pre_loop_end(cl);
154 if (pre_end == NULL) return;
155 Node *pre_opaq1 = pre_end->limit();
156 if (pre_opaq1->Opcode() != Op_Opaque1) return;
157 }
158
159 init(); // initialize data structures
160
161 set_lpt(lpt);
162 set_lp(cl);
163
164 // For now, define one block which is the entire loop body
165 set_bb(cl);
166
167 if (do_optimization) {
168 assert(_packset.length() == 0, "packset must be empty");
169 SLP_extract();
170 if (PostLoopMultiversioning && Matcher::has_predicated_vectors()) {
171 if (cl->is_vectorized_loop() && cl->is_main_loop() && !cl->is_reduction_loop()) {
172 IdealLoopTree *lpt_next = lpt->_next;
173 CountedLoopNode *cl_next = lpt_next->_head->as_CountedLoop();
174 _phase->has_range_checks(lpt_next);
175 if (cl_next->is_post_loop() && !cl_next->range_checks_present()) {
176 if (!cl_next->is_vectorized_loop()) {
177 int slp_max_unroll_factor = cl->slp_max_unroll();
178 cl_next->set_slp_max_unroll(slp_max_unroll_factor);
179 }
180 }
181 }
182 }
183 }
184 }
185
186 //------------------------------early unrolling analysis------------------------------
unrolling_analysis(int & local_loop_unroll_factor)187 void SuperWord::unrolling_analysis(int &local_loop_unroll_factor) {
188 bool is_slp = true;
189 ResourceMark rm;
190 size_t ignored_size = lpt()->_body.size();
191 int *ignored_loop_nodes = NEW_RESOURCE_ARRAY(int, ignored_size);
192 Node_Stack nstack((int)ignored_size);
193 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
194 Node *cl_exit = cl->loopexit_or_null();
195 int rpo_idx = _post_block.length();
196
197 assert(rpo_idx == 0, "post loop block is empty");
198
199 // First clear the entries
200 for (uint i = 0; i < lpt()->_body.size(); i++) {
201 ignored_loop_nodes[i] = -1;
202 }
203
204 int max_vector = Matcher::max_vector_size(T_BYTE);
205 bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
206
207 // Process the loop, some/all of the stack entries will not be in order, ergo
208 // need to preprocess the ignored initial state before we process the loop
209 for (uint i = 0; i < lpt()->_body.size(); i++) {
210 Node* n = lpt()->_body.at(i);
211 if (n == cl->incr() ||
212 n->is_reduction() ||
213 n->is_AddP() ||
214 n->is_Cmp() ||
215 n->is_IfTrue() ||
216 n->is_CountedLoop() ||
217 (n == cl_exit)) {
218 ignored_loop_nodes[i] = n->_idx;
219 continue;
220 }
221
222 if (n->is_If()) {
223 IfNode *iff = n->as_If();
224 if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
225 if (lpt()->is_loop_exit(iff)) {
226 ignored_loop_nodes[i] = n->_idx;
227 continue;
228 }
229 }
230 }
231
232 if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) {
233 Node* n_tail = n->in(LoopNode::LoopBackControl);
234 if (n_tail != n->in(LoopNode::EntryControl)) {
235 if (!n_tail->is_Mem()) {
236 is_slp = false;
237 break;
238 }
239 }
240 }
241
242 // This must happen after check of phi/if
243 if (n->is_Phi() || n->is_If()) {
244 ignored_loop_nodes[i] = n->_idx;
245 continue;
246 }
247
248 if (n->is_LoadStore() || n->is_MergeMem() ||
249 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
250 is_slp = false;
251 break;
252 }
253
254 // Ignore nodes with non-primitive type.
255 BasicType bt;
256 if (n->is_Mem()) {
257 bt = n->as_Mem()->memory_type();
258 } else {
259 bt = n->bottom_type()->basic_type();
260 }
261 if (is_java_primitive(bt) == false) {
262 ignored_loop_nodes[i] = n->_idx;
263 continue;
264 }
265
266 if (n->is_Mem()) {
267 MemNode* current = n->as_Mem();
268 Node* adr = n->in(MemNode::Address);
269 Node* n_ctrl = _phase->get_ctrl(adr);
270
271 // save a queue of post process nodes
272 if (n_ctrl != NULL && lpt()->is_member(_phase->get_loop(n_ctrl))) {
273 // Process the memory expression
274 int stack_idx = 0;
275 bool have_side_effects = true;
276 if (adr->is_AddP() == false) {
277 nstack.push(adr, stack_idx++);
278 } else {
279 // Mark the components of the memory operation in nstack
280 SWPointer p1(current, this, &nstack, true);
281 have_side_effects = p1.node_stack()->is_nonempty();
282 }
283
284 // Process the pointer stack
285 while (have_side_effects) {
286 Node* pointer_node = nstack.node();
287 for (uint j = 0; j < lpt()->_body.size(); j++) {
288 Node* cur_node = lpt()->_body.at(j);
289 if (cur_node == pointer_node) {
290 ignored_loop_nodes[j] = cur_node->_idx;
291 break;
292 }
293 }
294 nstack.pop();
295 have_side_effects = nstack.is_nonempty();
296 }
297 }
298 }
299 }
300
301 if (is_slp) {
302 // Now we try to find the maximum supported consistent vector which the machine
303 // description can use
304 bool small_basic_type = false;
305 bool flag_small_bt = false;
306 for (uint i = 0; i < lpt()->_body.size(); i++) {
307 if (ignored_loop_nodes[i] != -1) continue;
308
309 BasicType bt;
310 Node* n = lpt()->_body.at(i);
311 if (n->is_Mem()) {
312 bt = n->as_Mem()->memory_type();
313 } else {
314 bt = n->bottom_type()->basic_type();
315 }
316
317 if (post_loop_allowed) {
318 if (!small_basic_type) {
319 switch (bt) {
320 case T_CHAR:
321 case T_BYTE:
322 case T_SHORT:
323 small_basic_type = true;
324 break;
325
326 case T_LONG:
327 // TODO: Remove when support completed for mask context with LONG.
328 // Support needs to be augmented for logical qword operations, currently we map to dword
329 // buckets for vectors on logicals as these were legacy.
330 small_basic_type = true;
331 break;
332
333 default:
334 break;
335 }
336 }
337 }
338
339 if (is_java_primitive(bt) == false) continue;
340
341 int cur_max_vector = Matcher::max_vector_size(bt);
342
343 // If a max vector exists which is not larger than _local_loop_unroll_factor
344 // stop looking, we already have the max vector to map to.
345 if (cur_max_vector < local_loop_unroll_factor) {
346 is_slp = false;
347 if (TraceSuperWordLoopUnrollAnalysis) {
348 tty->print_cr("slp analysis fails: unroll limit greater than max vector\n");
349 }
350 break;
351 }
352
353 // Map the maximal common vector
354 if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) {
355 if (cur_max_vector < max_vector && !flag_small_bt) {
356 max_vector = cur_max_vector;
357 } else if (cur_max_vector > max_vector && UseSubwordForMaxVector) {
358 // Analyse subword in the loop to set maximum vector size to take advantage of full vector width for subword types.
359 // Here we analyze if narrowing is likely to happen and if it is we set vector size more aggressively.
360 // We check for possibility of narrowing by looking through chain operations using subword types.
361 if (is_subword_type(bt)) {
362 uint start, end;
363 VectorNode::vector_operands(n, &start, &end);
364
365 for (uint j = start; j < end; j++) {
366 Node* in = n->in(j);
367 // Don't propagate through a memory
368 if (!in->is_Mem() && in_bb(in) && in->bottom_type()->basic_type() == T_INT) {
369 bool same_type = true;
370 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
371 Node *use = in->fast_out(k);
372 if (!in_bb(use) && use->bottom_type()->basic_type() != bt) {
373 same_type = false;
374 break;
375 }
376 }
377 if (same_type) {
378 max_vector = cur_max_vector;
379 flag_small_bt = true;
380 cl->mark_subword_loop();
381 }
382 }
383 }
384 }
385 }
386 // We only process post loops on predicated targets where we want to
387 // mask map the loop to a single iteration
388 if (post_loop_allowed) {
389 _post_block.at_put_grow(rpo_idx++, n);
390 }
391 }
392 }
393 if (is_slp) {
394 local_loop_unroll_factor = max_vector;
395 cl->mark_passed_slp();
396 }
397 cl->mark_was_slp();
398 if (cl->is_main_loop()) {
399 cl->set_slp_max_unroll(local_loop_unroll_factor);
400 } else if (post_loop_allowed) {
401 if (!small_basic_type) {
402 // avoid replication context for small basic types in programmable masked loops
403 cl->set_slp_max_unroll(local_loop_unroll_factor);
404 }
405 }
406 }
407 }
408
409 //------------------------------SLP_extract---------------------------
410 // Extract the superword level parallelism
411 //
412 // 1) A reverse post-order of nodes in the block is constructed. By scanning
413 // this list from first to last, all definitions are visited before their uses.
414 //
415 // 2) A point-to-point dependence graph is constructed between memory references.
416 // This simplies the upcoming "independence" checker.
417 //
418 // 3) The maximum depth in the node graph from the beginning of the block
419 // to each node is computed. This is used to prune the graph search
420 // in the independence checker.
421 //
422 // 4) For integer types, the necessary bit width is propagated backwards
423 // from stores to allow packed operations on byte, char, and short
424 // integers. This reverses the promotion to type "int" that javac
425 // did for operations like: char c1,c2,c3; c1 = c2 + c3.
426 //
427 // 5) One of the memory references is picked to be an aligned vector reference.
428 // The pre-loop trip count is adjusted to align this reference in the
429 // unrolled body.
430 //
431 // 6) The initial set of pack pairs is seeded with memory references.
432 //
433 // 7) The set of pack pairs is extended by following use->def and def->use links.
434 //
435 // 8) The pairs are combined into vector sized packs.
436 //
437 // 9) Reorder the memory slices to co-locate members of the memory packs.
438 //
439 // 10) Generate ideal vector nodes for the final set of packs and where necessary,
440 // inserting scalar promotion, vector creation from multiple scalars, and
441 // extraction of scalar values from vectors.
442 //
SLP_extract()443 void SuperWord::SLP_extract() {
444
445 #ifndef PRODUCT
446 if (_do_vector_loop && TraceSuperWord) {
447 tty->print("SuperWord::SLP_extract\n");
448 tty->print("input loop\n");
449 _lpt->dump_head();
450 _lpt->dump();
451 for (uint i = 0; i < _lpt->_body.size(); i++) {
452 _lpt->_body.at(i)->dump();
453 }
454 }
455 #endif
456 // Ready the block
457 if (!construct_bb()) {
458 return; // Exit if no interesting nodes or complex graph.
459 }
460
461 // build _dg, _disjoint_ptrs
462 dependence_graph();
463
464 // compute function depth(Node*)
465 compute_max_depth();
466
467 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
468 bool post_loop_allowed = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
469 if (cl->is_main_loop()) {
470 if (_do_vector_loop) {
471 if (mark_generations() != -1) {
472 hoist_loads_in_graph(); // this only rebuild the graph; all basic structs need rebuild explicitly
473
474 if (!construct_bb()) {
475 return; // Exit if no interesting nodes or complex graph.
476 }
477 dependence_graph();
478 compute_max_depth();
479 }
480
481 #ifndef PRODUCT
482 if (TraceSuperWord) {
483 tty->print_cr("\nSuperWord::_do_vector_loop: graph after hoist_loads_in_graph");
484 _lpt->dump_head();
485 for (int j = 0; j < _block.length(); j++) {
486 Node* n = _block.at(j);
487 int d = depth(n);
488 for (int i = 0; i < d; i++) tty->print("%s", " ");
489 tty->print("%d :", d);
490 n->dump();
491 }
492 }
493 #endif
494 }
495
496 compute_vector_element_type();
497
498 // Attempt vectorization
499
500 find_adjacent_refs();
501
502 extend_packlist();
503
504 if (_do_vector_loop) {
505 if (_packset.length() == 0) {
506 if (TraceSuperWord) {
507 tty->print_cr("\nSuperWord::_do_vector_loop DFA could not build packset, now trying to build anyway");
508 }
509 pack_parallel();
510 }
511 }
512
513 combine_packs();
514
515 construct_my_pack_map();
516 if (UseVectorCmov) {
517 merge_packs_to_cmovd();
518 }
519
520 filter_packs();
521
522 schedule();
523 } else if (post_loop_allowed) {
524 int saved_mapped_unroll_factor = cl->slp_max_unroll();
525 if (saved_mapped_unroll_factor) {
526 int vector_mapped_unroll_factor = saved_mapped_unroll_factor;
527
528 // now reset the slp_unroll_factor so that we can check the analysis mapped
529 // what the vector loop was mapped to
530 cl->set_slp_max_unroll(0);
531
532 // do the analysis on the post loop
533 unrolling_analysis(vector_mapped_unroll_factor);
534
535 // if our analyzed loop is a canonical fit, start processing it
536 if (vector_mapped_unroll_factor == saved_mapped_unroll_factor) {
537 // now add the vector nodes to packsets
538 for (int i = 0; i < _post_block.length(); i++) {
539 Node* n = _post_block.at(i);
540 Node_List* singleton = new Node_List();
541 singleton->push(n);
542 _packset.append(singleton);
543 set_my_pack(n, singleton);
544 }
545
546 // map base types for vector usage
547 compute_vector_element_type();
548 } else {
549 return;
550 }
551 } else {
552 // for some reason we could not map the slp analysis state of the vectorized loop
553 return;
554 }
555 }
556
557 output();
558 }
559
560 //------------------------------find_adjacent_refs---------------------------
561 // Find the adjacent memory references and create pack pairs for them.
562 // This is the initial set of packs that will then be extended by
563 // following use->def and def->use links. The align positions are
564 // assigned relative to the reference "align_to_ref"
find_adjacent_refs()565 void SuperWord::find_adjacent_refs() {
566 // Get list of memory operations
567 Node_List memops;
568 for (int i = 0; i < _block.length(); i++) {
569 Node* n = _block.at(i);
570 if (n->is_Mem() && !n->is_LoadStore() && in_bb(n) &&
571 is_java_primitive(n->as_Mem()->memory_type())) {
572 int align = memory_alignment(n->as_Mem(), 0);
573 if (align != bottom_align) {
574 memops.push(n);
575 }
576 }
577 }
578
579 Node_List align_to_refs;
580 int max_idx;
581 int best_iv_adjustment = 0;
582 MemNode* best_align_to_mem_ref = NULL;
583
584 while (memops.size() != 0) {
585 // Find a memory reference to align to.
586 MemNode* mem_ref = find_align_to_ref(memops, max_idx);
587 if (mem_ref == NULL) break;
588 align_to_refs.push(mem_ref);
589 int iv_adjustment = get_iv_adjustment(mem_ref);
590
591 if (best_align_to_mem_ref == NULL) {
592 // Set memory reference which is the best from all memory operations
593 // to be used for alignment. The pre-loop trip count is modified to align
594 // this reference to a vector-aligned address.
595 best_align_to_mem_ref = mem_ref;
596 best_iv_adjustment = iv_adjustment;
597 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
598 }
599
600 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
601 // Set alignment relative to "align_to_ref" for all related memory operations.
602 for (int i = memops.size() - 1; i >= 0; i--) {
603 MemNode* s = memops.at(i)->as_Mem();
604 if (isomorphic(s, mem_ref) &&
605 (!_do_vector_loop || same_origin_idx(s, mem_ref))) {
606 SWPointer p2(s, this, NULL, false);
607 if (p2.comparable(align_to_ref_p)) {
608 int align = memory_alignment(s, iv_adjustment);
609 set_alignment(s, align);
610 }
611 }
612 }
613
614 // Create initial pack pairs of memory operations for which
615 // alignment is set and vectors will be aligned.
616 bool create_pack = true;
617 if (memory_alignment(mem_ref, best_iv_adjustment) == 0 || _do_vector_loop) {
618 if (vectors_should_be_aligned()) {
619 int vw = vector_width(mem_ref);
620 int vw_best = vector_width(best_align_to_mem_ref);
621 if (vw > vw_best) {
622 // Do not vectorize a memory access with more elements per vector
623 // if unaligned memory access is not allowed because number of
624 // iterations in pre-loop will be not enough to align it.
625 create_pack = false;
626 } else {
627 SWPointer p2(best_align_to_mem_ref, this, NULL, false);
628 if (align_to_ref_p.invar() != p2.invar()) {
629 // Do not vectorize memory accesses with different invariants
630 // if unaligned memory accesses are not allowed.
631 create_pack = false;
632 }
633 }
634 }
635 } else {
636 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
637 // Can't allow vectorization of unaligned memory accesses with the
638 // same type since it could be overlapped accesses to the same array.
639 create_pack = false;
640 } else {
641 // Allow independent (different type) unaligned memory operations
642 // if HW supports them.
643 if (vectors_should_be_aligned()) {
644 create_pack = false;
645 } else {
646 // Check if packs of the same memory type but
647 // with a different alignment were created before.
648 for (uint i = 0; i < align_to_refs.size(); i++) {
649 MemNode* mr = align_to_refs.at(i)->as_Mem();
650 if (mr == mem_ref) {
651 // Skip when we are looking at same memory operation.
652 continue;
653 }
654 if (same_velt_type(mr, mem_ref) &&
655 memory_alignment(mr, iv_adjustment) != 0)
656 create_pack = false;
657 }
658 }
659 }
660 }
661 if (create_pack) {
662 for (uint i = 0; i < memops.size(); i++) {
663 Node* s1 = memops.at(i);
664 int align = alignment(s1);
665 if (align == top_align) continue;
666 for (uint j = 0; j < memops.size(); j++) {
667 Node* s2 = memops.at(j);
668 if (alignment(s2) == top_align) continue;
669 if (s1 != s2 && are_adjacent_refs(s1, s2)) {
670 if (stmts_can_pack(s1, s2, align)) {
671 Node_List* pair = new Node_List();
672 pair->push(s1);
673 pair->push(s2);
674 if (!_do_vector_loop || same_origin_idx(s1, s2)) {
675 _packset.append(pair);
676 }
677 }
678 }
679 }
680 }
681 } else { // Don't create unaligned pack
682 // First, remove remaining memory ops of the same type from the list.
683 for (int i = memops.size() - 1; i >= 0; i--) {
684 MemNode* s = memops.at(i)->as_Mem();
685 if (same_velt_type(s, mem_ref)) {
686 memops.remove(i);
687 }
688 }
689
690 // Second, remove already constructed packs of the same type.
691 for (int i = _packset.length() - 1; i >= 0; i--) {
692 Node_List* p = _packset.at(i);
693 MemNode* s = p->at(0)->as_Mem();
694 if (same_velt_type(s, mem_ref)) {
695 remove_pack_at(i);
696 }
697 }
698
699 // If needed find the best memory reference for loop alignment again.
700 if (same_velt_type(mem_ref, best_align_to_mem_ref)) {
701 // Put memory ops from remaining packs back on memops list for
702 // the best alignment search.
703 uint orig_msize = memops.size();
704 for (int i = 0; i < _packset.length(); i++) {
705 Node_List* p = _packset.at(i);
706 MemNode* s = p->at(0)->as_Mem();
707 assert(!same_velt_type(s, mem_ref), "sanity");
708 memops.push(s);
709 }
710 best_align_to_mem_ref = find_align_to_ref(memops, max_idx);
711 if (best_align_to_mem_ref == NULL) {
712 if (TraceSuperWord) {
713 tty->print_cr("SuperWord::find_adjacent_refs(): best_align_to_mem_ref == NULL");
714 }
715 // best_align_to_mem_ref will be used for adjusting the pre-loop limit in
716 // SuperWord::align_initial_loop_index. Find one with the biggest vector size,
717 // smallest data size and smallest iv offset from memory ops from remaining packs.
718 if (_packset.length() > 0) {
719 if (orig_msize == 0) {
720 best_align_to_mem_ref = memops.at(max_idx)->as_Mem();
721 } else {
722 for (uint i = 0; i < orig_msize; i++) {
723 memops.remove(0);
724 }
725 best_align_to_mem_ref = find_align_to_ref(memops, max_idx);
726 assert(best_align_to_mem_ref == NULL, "sanity");
727 best_align_to_mem_ref = memops.at(max_idx)->as_Mem();
728 }
729 assert(best_align_to_mem_ref != NULL, "sanity");
730 }
731 break;
732 }
733 best_iv_adjustment = get_iv_adjustment(best_align_to_mem_ref);
734 NOT_PRODUCT(find_adjacent_refs_trace_1(best_align_to_mem_ref, best_iv_adjustment);)
735 // Restore list.
736 while (memops.size() > orig_msize)
737 (void)memops.pop();
738 }
739 } // unaligned memory accesses
740
741 // Remove used mem nodes.
742 for (int i = memops.size() - 1; i >= 0; i--) {
743 MemNode* m = memops.at(i)->as_Mem();
744 if (alignment(m) != top_align) {
745 memops.remove(i);
746 }
747 }
748
749 } // while (memops.size() != 0
750 set_align_to_ref(best_align_to_mem_ref);
751
752 if (TraceSuperWord) {
753 tty->print_cr("\nAfter find_adjacent_refs");
754 print_packset();
755 }
756 }
757
758 #ifndef PRODUCT
find_adjacent_refs_trace_1(Node * best_align_to_mem_ref,int best_iv_adjustment)759 void SuperWord::find_adjacent_refs_trace_1(Node* best_align_to_mem_ref, int best_iv_adjustment) {
760 if (is_trace_adjacent()) {
761 tty->print("SuperWord::find_adjacent_refs best_align_to_mem_ref = %d, best_iv_adjustment = %d",
762 best_align_to_mem_ref->_idx, best_iv_adjustment);
763 best_align_to_mem_ref->dump();
764 }
765 }
766 #endif
767
768 //------------------------------find_align_to_ref---------------------------
769 // Find a memory reference to align the loop induction variable to.
770 // Looks first at stores then at loads, looking for a memory reference
771 // with the largest number of references similar to it.
find_align_to_ref(Node_List & memops,int & idx)772 MemNode* SuperWord::find_align_to_ref(Node_List &memops, int &idx) {
773 GrowableArray<int> cmp_ct(arena(), memops.size(), memops.size(), 0);
774
775 // Count number of comparable memory ops
776 for (uint i = 0; i < memops.size(); i++) {
777 MemNode* s1 = memops.at(i)->as_Mem();
778 SWPointer p1(s1, this, NULL, false);
779 // Only discard unalignable memory references if vector memory references
780 // should be aligned on this platform.
781 if (vectors_should_be_aligned() && !ref_is_alignable(p1)) {
782 *cmp_ct.adr_at(i) = 0;
783 continue;
784 }
785 for (uint j = i+1; j < memops.size(); j++) {
786 MemNode* s2 = memops.at(j)->as_Mem();
787 if (isomorphic(s1, s2)) {
788 SWPointer p2(s2, this, NULL, false);
789 if (p1.comparable(p2)) {
790 (*cmp_ct.adr_at(i))++;
791 (*cmp_ct.adr_at(j))++;
792 }
793 }
794 }
795 }
796
797 // Find Store (or Load) with the greatest number of "comparable" references,
798 // biggest vector size, smallest data size and smallest iv offset.
799 int max_ct = 0;
800 int max_vw = 0;
801 int max_idx = -1;
802 int min_size = max_jint;
803 int min_iv_offset = max_jint;
804 for (uint j = 0; j < memops.size(); j++) {
805 MemNode* s = memops.at(j)->as_Mem();
806 if (s->is_Store()) {
807 int vw = vector_width_in_bytes(s);
808 assert(vw > 1, "sanity");
809 SWPointer p(s, this, NULL, false);
810 if ( cmp_ct.at(j) > max_ct ||
811 (cmp_ct.at(j) == max_ct &&
812 ( vw > max_vw ||
813 (vw == max_vw &&
814 ( data_size(s) < min_size ||
815 (data_size(s) == min_size &&
816 p.offset_in_bytes() < min_iv_offset)))))) {
817 max_ct = cmp_ct.at(j);
818 max_vw = vw;
819 max_idx = j;
820 min_size = data_size(s);
821 min_iv_offset = p.offset_in_bytes();
822 }
823 }
824 }
825 // If no stores, look at loads
826 if (max_ct == 0) {
827 for (uint j = 0; j < memops.size(); j++) {
828 MemNode* s = memops.at(j)->as_Mem();
829 if (s->is_Load()) {
830 int vw = vector_width_in_bytes(s);
831 assert(vw > 1, "sanity");
832 SWPointer p(s, this, NULL, false);
833 if ( cmp_ct.at(j) > max_ct ||
834 (cmp_ct.at(j) == max_ct &&
835 ( vw > max_vw ||
836 (vw == max_vw &&
837 ( data_size(s) < min_size ||
838 (data_size(s) == min_size &&
839 p.offset_in_bytes() < min_iv_offset)))))) {
840 max_ct = cmp_ct.at(j);
841 max_vw = vw;
842 max_idx = j;
843 min_size = data_size(s);
844 min_iv_offset = p.offset_in_bytes();
845 }
846 }
847 }
848 }
849
850 #ifdef ASSERT
851 if (TraceSuperWord && Verbose) {
852 tty->print_cr("\nVector memops after find_align_to_ref");
853 for (uint i = 0; i < memops.size(); i++) {
854 MemNode* s = memops.at(i)->as_Mem();
855 s->dump();
856 }
857 }
858 #endif
859
860 idx = max_idx;
861 if (max_ct > 0) {
862 #ifdef ASSERT
863 if (TraceSuperWord) {
864 tty->print("\nVector align to node: ");
865 memops.at(max_idx)->as_Mem()->dump();
866 }
867 #endif
868 return memops.at(max_idx)->as_Mem();
869 }
870 return NULL;
871 }
872
873 //------------------span_works_for_memory_size-----------------------------
span_works_for_memory_size(MemNode * mem,int span,int mem_size,int offset)874 static bool span_works_for_memory_size(MemNode* mem, int span, int mem_size, int offset) {
875 bool span_matches_memory = false;
876 if ((mem_size == type2aelembytes(T_BYTE) || mem_size == type2aelembytes(T_SHORT))
877 && ABS(span) == type2aelembytes(T_INT)) {
878 // There is a mismatch on span size compared to memory.
879 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
880 Node* use = mem->fast_out(j);
881 if (!VectorNode::is_type_transition_to_int(use)) {
882 return false;
883 }
884 }
885 // If all uses transition to integer, it means that we can successfully align even on mismatch.
886 return true;
887 }
888 else {
889 span_matches_memory = ABS(span) == mem_size;
890 }
891 return span_matches_memory && (ABS(offset) % mem_size) == 0;
892 }
893
894 //------------------------------ref_is_alignable---------------------------
895 // Can the preloop align the reference to position zero in the vector?
ref_is_alignable(SWPointer & p)896 bool SuperWord::ref_is_alignable(SWPointer& p) {
897 if (!p.has_iv()) {
898 return true; // no induction variable
899 }
900 CountedLoopEndNode* pre_end = get_pre_loop_end(lp()->as_CountedLoop());
901 assert(pre_end != NULL, "we must have a correct pre-loop");
902 assert(pre_end->stride_is_con(), "pre loop stride is constant");
903 int preloop_stride = pre_end->stride_con();
904
905 int span = preloop_stride * p.scale_in_bytes();
906 int mem_size = p.memory_size();
907 int offset = p.offset_in_bytes();
908 // Stride one accesses are alignable if offset is aligned to memory operation size.
909 // Offset can be unaligned when UseUnalignedAccesses is used.
910 if (span_works_for_memory_size(p.mem(), span, mem_size, offset)) {
911 return true;
912 }
913 // If the initial offset from start of the object is computable,
914 // check if the pre-loop can align the final offset accordingly.
915 //
916 // In other words: Can we find an i such that the offset
917 // after i pre-loop iterations is aligned to vw?
918 // (init_offset + pre_loop) % vw == 0 (1)
919 // where
920 // pre_loop = i * span
921 // is the number of bytes added to the offset by i pre-loop iterations.
922 //
923 // For this to hold we need pre_loop to increase init_offset by
924 // pre_loop = vw - (init_offset % vw)
925 //
926 // This is only possible if pre_loop is divisible by span because each
927 // pre-loop iteration increases the initial offset by 'span' bytes:
928 // (vw - (init_offset % vw)) % span == 0
929 //
930 int vw = vector_width_in_bytes(p.mem());
931 assert(vw > 1, "sanity");
932 Node* init_nd = pre_end->init_trip();
933 if (init_nd->is_Con() && p.invar() == NULL) {
934 int init = init_nd->bottom_type()->is_int()->get_con();
935 int init_offset = init * p.scale_in_bytes() + offset;
936 if (init_offset < 0) { // negative offset from object start?
937 return false; // may happen in dead loop
938 }
939 if (vw % span == 0) {
940 // If vm is a multiple of span, we use formula (1).
941 if (span > 0) {
942 return (vw - (init_offset % vw)) % span == 0;
943 } else {
944 assert(span < 0, "nonzero stride * scale");
945 return (init_offset % vw) % -span == 0;
946 }
947 } else if (span % vw == 0) {
948 // If span is a multiple of vw, we can simplify formula (1) to:
949 // (init_offset + i * span) % vw == 0
950 // =>
951 // (init_offset % vw) + ((i * span) % vw) == 0
952 // =>
953 // init_offset % vw == 0
954 //
955 // Because we add a multiple of vw to the initial offset, the final
956 // offset is a multiple of vw if and only if init_offset is a multiple.
957 //
958 return (init_offset % vw) == 0;
959 }
960 }
961 return false;
962 }
963 //---------------------------get_vw_bytes_special------------------------
get_vw_bytes_special(MemNode * s)964 int SuperWord::get_vw_bytes_special(MemNode* s) {
965 // Get the vector width in bytes.
966 int vw = vector_width_in_bytes(s);
967
968 // Check for special case where there is an MulAddS2I usage where short vectors are going to need combined.
969 BasicType btype = velt_basic_type(s);
970 if (type2aelembytes(btype) == 2) {
971 bool should_combine_adjacent = true;
972 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
973 Node* user = s->fast_out(i);
974 if (!VectorNode::is_muladds2i(user)) {
975 should_combine_adjacent = false;
976 }
977 }
978 if (should_combine_adjacent) {
979 vw = MIN2(Matcher::max_vector_size(btype)*type2aelembytes(btype), vw * 2);
980 }
981 }
982
983 return vw;
984 }
985
986 //---------------------------get_iv_adjustment---------------------------
987 // Calculate loop's iv adjustment for this memory ops.
get_iv_adjustment(MemNode * mem_ref)988 int SuperWord::get_iv_adjustment(MemNode* mem_ref) {
989 SWPointer align_to_ref_p(mem_ref, this, NULL, false);
990 int offset = align_to_ref_p.offset_in_bytes();
991 int scale = align_to_ref_p.scale_in_bytes();
992 int elt_size = align_to_ref_p.memory_size();
993 int vw = get_vw_bytes_special(mem_ref);
994 assert(vw > 1, "sanity");
995 int iv_adjustment;
996 if (scale != 0) {
997 int stride_sign = (scale * iv_stride()) > 0 ? 1 : -1;
998 // At least one iteration is executed in pre-loop by default. As result
999 // several iterations are needed to align memory operations in main-loop even
1000 // if offset is 0.
1001 int iv_adjustment_in_bytes = (stride_sign * vw - (offset % vw));
1002 // iv_adjustment_in_bytes must be a multiple of elt_size if vector memory
1003 // references should be aligned on this platform.
1004 assert((ABS(iv_adjustment_in_bytes) % elt_size) == 0 || !vectors_should_be_aligned(),
1005 "(%d) should be divisible by (%d)", iv_adjustment_in_bytes, elt_size);
1006 iv_adjustment = iv_adjustment_in_bytes/elt_size;
1007 } else {
1008 // This memory op is not dependent on iv (scale == 0)
1009 iv_adjustment = 0;
1010 }
1011
1012 #ifndef PRODUCT
1013 if (TraceSuperWord) {
1014 tty->print("SuperWord::get_iv_adjustment: n = %d, noffset = %d iv_adjust = %d elt_size = %d scale = %d iv_stride = %d vect_size %d: ",
1015 mem_ref->_idx, offset, iv_adjustment, elt_size, scale, iv_stride(), vw);
1016 mem_ref->dump();
1017 }
1018 #endif
1019 return iv_adjustment;
1020 }
1021
1022 //---------------------------dependence_graph---------------------------
1023 // Construct dependency graph.
1024 // Add dependence edges to load/store nodes for memory dependence
1025 // A.out()->DependNode.in(1) and DependNode.out()->B.prec(x)
dependence_graph()1026 void SuperWord::dependence_graph() {
1027 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
1028 // First, assign a dependence node to each memory node
1029 for (int i = 0; i < _block.length(); i++ ) {
1030 Node *n = _block.at(i);
1031 if (n->is_Mem() || (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
1032 _dg.make_node(n);
1033 }
1034 }
1035
1036 // For each memory slice, create the dependences
1037 for (int i = 0; i < _mem_slice_head.length(); i++) {
1038 Node* n = _mem_slice_head.at(i);
1039 Node* n_tail = _mem_slice_tail.at(i);
1040
1041 // Get slice in predecessor order (last is first)
1042 if (cl->is_main_loop()) {
1043 mem_slice_preds(n_tail, n, _nlist);
1044 }
1045
1046 #ifndef PRODUCT
1047 if(TraceSuperWord && Verbose) {
1048 tty->print_cr("SuperWord::dependence_graph: built a new mem slice");
1049 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
1050 _nlist.at(j)->dump();
1051 }
1052 }
1053 #endif
1054 // Make the slice dependent on the root
1055 DepMem* slice = _dg.dep(n);
1056 _dg.make_edge(_dg.root(), slice);
1057
1058 // Create a sink for the slice
1059 DepMem* slice_sink = _dg.make_node(NULL);
1060 _dg.make_edge(slice_sink, _dg.tail());
1061
1062 // Now visit each pair of memory ops, creating the edges
1063 for (int j = _nlist.length() - 1; j >= 0 ; j--) {
1064 Node* s1 = _nlist.at(j);
1065
1066 // If no dependency yet, use slice
1067 if (_dg.dep(s1)->in_cnt() == 0) {
1068 _dg.make_edge(slice, s1);
1069 }
1070 SWPointer p1(s1->as_Mem(), this, NULL, false);
1071 bool sink_dependent = true;
1072 for (int k = j - 1; k >= 0; k--) {
1073 Node* s2 = _nlist.at(k);
1074 if (s1->is_Load() && s2->is_Load())
1075 continue;
1076 SWPointer p2(s2->as_Mem(), this, NULL, false);
1077
1078 int cmp = p1.cmp(p2);
1079 if (SuperWordRTDepCheck &&
1080 p1.base() != p2.base() && p1.valid() && p2.valid()) {
1081 // Create a runtime check to disambiguate
1082 OrderedPair pp(p1.base(), p2.base());
1083 _disjoint_ptrs.append_if_missing(pp);
1084 } else if (!SWPointer::not_equal(cmp)) {
1085 // Possibly same address
1086 _dg.make_edge(s1, s2);
1087 sink_dependent = false;
1088 }
1089 }
1090 if (sink_dependent) {
1091 _dg.make_edge(s1, slice_sink);
1092 }
1093 }
1094
1095 if (TraceSuperWord) {
1096 tty->print_cr("\nDependence graph for slice: %d", n->_idx);
1097 for (int q = 0; q < _nlist.length(); q++) {
1098 _dg.print(_nlist.at(q));
1099 }
1100 tty->cr();
1101 }
1102
1103 _nlist.clear();
1104 }
1105
1106 if (TraceSuperWord) {
1107 tty->print_cr("\ndisjoint_ptrs: %s", _disjoint_ptrs.length() > 0 ? "" : "NONE");
1108 for (int r = 0; r < _disjoint_ptrs.length(); r++) {
1109 _disjoint_ptrs.at(r).print();
1110 tty->cr();
1111 }
1112 tty->cr();
1113 }
1114
1115 }
1116
1117 //---------------------------mem_slice_preds---------------------------
1118 // Return a memory slice (node list) in predecessor order starting at "start"
mem_slice_preds(Node * start,Node * stop,GrowableArray<Node * > & preds)1119 void SuperWord::mem_slice_preds(Node* start, Node* stop, GrowableArray<Node*> &preds) {
1120 assert(preds.length() == 0, "start empty");
1121 Node* n = start;
1122 Node* prev = NULL;
1123 while (true) {
1124 NOT_PRODUCT( if(is_trace_mem_slice()) tty->print_cr("SuperWord::mem_slice_preds: n %d", n->_idx);)
1125 assert(in_bb(n), "must be in block");
1126 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1127 Node* out = n->fast_out(i);
1128 if (out->is_Load()) {
1129 if (in_bb(out)) {
1130 preds.push(out);
1131 if (TraceSuperWord && Verbose) {
1132 tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", out->_idx);
1133 }
1134 }
1135 } else {
1136 // FIXME
1137 if (out->is_MergeMem() && !in_bb(out)) {
1138 // Either unrolling is causing a memory edge not to disappear,
1139 // or need to run igvn.optimize() again before SLP
1140 } else if (out->is_Phi() && out->bottom_type() == Type::MEMORY && !in_bb(out)) {
1141 // Ditto. Not sure what else to check further.
1142 } else if (out->Opcode() == Op_StoreCM && out->in(MemNode::OopStore) == n) {
1143 // StoreCM has an input edge used as a precedence edge.
1144 // Maybe an issue when oop stores are vectorized.
1145 } else {
1146 assert(out == prev || prev == NULL, "no branches off of store slice");
1147 }
1148 }//else
1149 }//for
1150 if (n == stop) break;
1151 preds.push(n);
1152 if (TraceSuperWord && Verbose) {
1153 tty->print_cr("SuperWord::mem_slice_preds: added pred(%d)", n->_idx);
1154 }
1155 prev = n;
1156 assert(n->is_Mem(), "unexpected node %s", n->Name());
1157 n = n->in(MemNode::Memory);
1158 }
1159 }
1160
1161 //------------------------------stmts_can_pack---------------------------
1162 // Can s1 and s2 be in a pack with s1 immediately preceding s2 and
1163 // s1 aligned at "align"
stmts_can_pack(Node * s1,Node * s2,int align)1164 bool SuperWord::stmts_can_pack(Node* s1, Node* s2, int align) {
1165
1166 // Do not use superword for non-primitives
1167 BasicType bt1 = velt_basic_type(s1);
1168 BasicType bt2 = velt_basic_type(s2);
1169 if(!is_java_primitive(bt1) || !is_java_primitive(bt2))
1170 return false;
1171 if (Matcher::max_vector_size(bt1) < 2) {
1172 return false; // No vectors for this type
1173 }
1174
1175 if (isomorphic(s1, s2)) {
1176 if ((independent(s1, s2) && have_similar_inputs(s1, s2)) || reduction(s1, s2)) {
1177 if (!exists_at(s1, 0) && !exists_at(s2, 1)) {
1178 if (!s1->is_Mem() || are_adjacent_refs(s1, s2)) {
1179 int s1_align = alignment(s1);
1180 int s2_align = alignment(s2);
1181 if (s1_align == top_align || s1_align == align) {
1182 if (s2_align == top_align || s2_align == align + data_size(s1)) {
1183 return true;
1184 }
1185 }
1186 }
1187 }
1188 }
1189 }
1190 return false;
1191 }
1192
1193 //------------------------------exists_at---------------------------
1194 // Does s exist in a pack at position pos?
exists_at(Node * s,uint pos)1195 bool SuperWord::exists_at(Node* s, uint pos) {
1196 for (int i = 0; i < _packset.length(); i++) {
1197 Node_List* p = _packset.at(i);
1198 if (p->at(pos) == s) {
1199 return true;
1200 }
1201 }
1202 return false;
1203 }
1204
1205 //------------------------------are_adjacent_refs---------------------------
1206 // Is s1 immediately before s2 in memory?
are_adjacent_refs(Node * s1,Node * s2)1207 bool SuperWord::are_adjacent_refs(Node* s1, Node* s2) {
1208 if (!s1->is_Mem() || !s2->is_Mem()) return false;
1209 if (!in_bb(s1) || !in_bb(s2)) return false;
1210
1211 // Do not use superword for non-primitives
1212 if (!is_java_primitive(s1->as_Mem()->memory_type()) ||
1213 !is_java_primitive(s2->as_Mem()->memory_type())) {
1214 return false;
1215 }
1216
1217 // FIXME - co_locate_pack fails on Stores in different mem-slices, so
1218 // only pack memops that are in the same alias set until that's fixed.
1219 if (_phase->C->get_alias_index(s1->as_Mem()->adr_type()) !=
1220 _phase->C->get_alias_index(s2->as_Mem()->adr_type()))
1221 return false;
1222 SWPointer p1(s1->as_Mem(), this, NULL, false);
1223 SWPointer p2(s2->as_Mem(), this, NULL, false);
1224 if (p1.base() != p2.base() || !p1.comparable(p2)) return false;
1225 int diff = p2.offset_in_bytes() - p1.offset_in_bytes();
1226 return diff == data_size(s1);
1227 }
1228
1229 //------------------------------isomorphic---------------------------
1230 // Are s1 and s2 similar?
isomorphic(Node * s1,Node * s2)1231 bool SuperWord::isomorphic(Node* s1, Node* s2) {
1232 if (s1->Opcode() != s2->Opcode()) return false;
1233 if (s1->req() != s2->req()) return false;
1234 if (!same_velt_type(s1, s2)) return false;
1235 Node* s1_ctrl = s1->in(0);
1236 Node* s2_ctrl = s2->in(0);
1237 // If the control nodes are equivalent, no further checks are required to test for isomorphism.
1238 if (s1_ctrl == s2_ctrl) {
1239 return true;
1240 } else {
1241 bool s1_ctrl_inv = ((s1_ctrl == NULL) ? true : lpt()->is_invariant(s1_ctrl));
1242 bool s2_ctrl_inv = ((s2_ctrl == NULL) ? true : lpt()->is_invariant(s2_ctrl));
1243 // If the control nodes are not invariant for the loop, fail isomorphism test.
1244 if (!s1_ctrl_inv || !s2_ctrl_inv) {
1245 return false;
1246 }
1247 if(s1_ctrl != NULL && s2_ctrl != NULL) {
1248 if (s1_ctrl->is_Proj()) {
1249 s1_ctrl = s1_ctrl->in(0);
1250 assert(lpt()->is_invariant(s1_ctrl), "must be invariant");
1251 }
1252 if (s2_ctrl->is_Proj()) {
1253 s2_ctrl = s2_ctrl->in(0);
1254 assert(lpt()->is_invariant(s2_ctrl), "must be invariant");
1255 }
1256 if (!s1_ctrl->is_RangeCheck() || !s2_ctrl->is_RangeCheck()) {
1257 return false;
1258 }
1259 }
1260 // Control nodes are invariant. However, we have no way of checking whether they resolve
1261 // in an equivalent manner. But, we know that invariant range checks are guaranteed to
1262 // throw before the loop (if they would have thrown). Thus, the loop would not have been reached.
1263 // Therefore, if the control nodes for both are range checks, we accept them to be isomorphic.
1264 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1265 Node* t1 = s1->fast_out(i);
1266 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1267 Node* t2 = s2->fast_out(j);
1268 if (VectorNode::is_muladds2i(t1) && VectorNode::is_muladds2i(t2)) {
1269 return true;
1270 }
1271 }
1272 }
1273 }
1274 return false;
1275 }
1276
1277 //------------------------------independent---------------------------
1278 // Is there no data path from s1 to s2 or s2 to s1?
independent(Node * s1,Node * s2)1279 bool SuperWord::independent(Node* s1, Node* s2) {
1280 // assert(s1->Opcode() == s2->Opcode(), "check isomorphic first");
1281 int d1 = depth(s1);
1282 int d2 = depth(s2);
1283 if (d1 == d2) return s1 != s2;
1284 Node* deep = d1 > d2 ? s1 : s2;
1285 Node* shallow = d1 > d2 ? s2 : s1;
1286
1287 visited_clear();
1288
1289 return independent_path(shallow, deep);
1290 }
1291
1292 //--------------------------have_similar_inputs-----------------------
1293 // For a node pair (s1, s2) which is isomorphic and independent,
1294 // do s1 and s2 have similar input edges?
have_similar_inputs(Node * s1,Node * s2)1295 bool SuperWord::have_similar_inputs(Node* s1, Node* s2) {
1296 // assert(isomorphic(s1, s2) == true, "check isomorphic");
1297 // assert(independent(s1, s2) == true, "check independent");
1298 if (s1->req() > 1 && !s1->is_Store() && !s1->is_Load()) {
1299 for (uint i = 1; i < s1->req(); i++) {
1300 if (s1->in(i)->Opcode() != s2->in(i)->Opcode()) return false;
1301 }
1302 }
1303 return true;
1304 }
1305
1306 //------------------------------reduction---------------------------
1307 // Is there a data path between s1 and s2 and the nodes reductions?
reduction(Node * s1,Node * s2)1308 bool SuperWord::reduction(Node* s1, Node* s2) {
1309 bool retValue = false;
1310 int d1 = depth(s1);
1311 int d2 = depth(s2);
1312 if (d1 + 1 == d2) {
1313 if (s1->is_reduction() && s2->is_reduction()) {
1314 // This is an ordered set, so s1 should define s2
1315 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1316 Node* t1 = s1->fast_out(i);
1317 if (t1 == s2) {
1318 // both nodes are reductions and connected
1319 retValue = true;
1320 }
1321 }
1322 }
1323 }
1324
1325 return retValue;
1326 }
1327
1328 //------------------------------independent_path------------------------------
1329 // Helper for independent
independent_path(Node * shallow,Node * deep,uint dp)1330 bool SuperWord::independent_path(Node* shallow, Node* deep, uint dp) {
1331 if (dp >= 1000) return false; // stop deep recursion
1332 visited_set(deep);
1333 int shal_depth = depth(shallow);
1334 assert(shal_depth <= depth(deep), "must be");
1335 for (DepPreds preds(deep, _dg); !preds.done(); preds.next()) {
1336 Node* pred = preds.current();
1337 if (in_bb(pred) && !visited_test(pred)) {
1338 if (shallow == pred) {
1339 return false;
1340 }
1341 if (shal_depth < depth(pred) && !independent_path(shallow, pred, dp+1)) {
1342 return false;
1343 }
1344 }
1345 }
1346 return true;
1347 }
1348
1349 //------------------------------set_alignment---------------------------
set_alignment(Node * s1,Node * s2,int align)1350 void SuperWord::set_alignment(Node* s1, Node* s2, int align) {
1351 set_alignment(s1, align);
1352 if (align == top_align || align == bottom_align) {
1353 set_alignment(s2, align);
1354 } else {
1355 set_alignment(s2, align + data_size(s1));
1356 }
1357 }
1358
1359 //------------------------------data_size---------------------------
data_size(Node * s)1360 int SuperWord::data_size(Node* s) {
1361 Node* use = NULL; //test if the node is a candidate for CMoveV optimization, then return the size of CMov
1362 if (UseVectorCmov) {
1363 use = _cmovev_kit.is_Bool_candidate(s);
1364 if (use != NULL) {
1365 return data_size(use);
1366 }
1367 use = _cmovev_kit.is_CmpD_candidate(s);
1368 if (use != NULL) {
1369 return data_size(use);
1370 }
1371 }
1372
1373 int bsize = type2aelembytes(velt_basic_type(s));
1374 assert(bsize != 0, "valid size");
1375 return bsize;
1376 }
1377
1378 //------------------------------extend_packlist---------------------------
1379 // Extend packset by following use->def and def->use links from pack members.
extend_packlist()1380 void SuperWord::extend_packlist() {
1381 bool changed;
1382 do {
1383 packset_sort(_packset.length());
1384 changed = false;
1385 for (int i = 0; i < _packset.length(); i++) {
1386 Node_List* p = _packset.at(i);
1387 changed |= follow_use_defs(p);
1388 changed |= follow_def_uses(p);
1389 }
1390 } while (changed);
1391
1392 if (_race_possible) {
1393 for (int i = 0; i < _packset.length(); i++) {
1394 Node_List* p = _packset.at(i);
1395 order_def_uses(p);
1396 }
1397 }
1398
1399 if (TraceSuperWord) {
1400 tty->print_cr("\nAfter extend_packlist");
1401 print_packset();
1402 }
1403 }
1404
1405 //------------------------------follow_use_defs---------------------------
1406 // Extend the packset by visiting operand definitions of nodes in pack p
follow_use_defs(Node_List * p)1407 bool SuperWord::follow_use_defs(Node_List* p) {
1408 assert(p->size() == 2, "just checking");
1409 Node* s1 = p->at(0);
1410 Node* s2 = p->at(1);
1411 assert(s1->req() == s2->req(), "just checking");
1412 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1413
1414 if (s1->is_Load()) return false;
1415
1416 int align = alignment(s1);
1417 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: s1 %d, align %d", s1->_idx, align);)
1418 bool changed = false;
1419 int start = s1->is_Store() ? MemNode::ValueIn : 1;
1420 int end = s1->is_Store() ? MemNode::ValueIn+1 : s1->req();
1421 for (int j = start; j < end; j++) {
1422 Node* t1 = s1->in(j);
1423 Node* t2 = s2->in(j);
1424 if (!in_bb(t1) || !in_bb(t2))
1425 continue;
1426 if (stmts_can_pack(t1, t2, align)) {
1427 if (est_savings(t1, t2) >= 0) {
1428 Node_List* pair = new Node_List();
1429 pair->push(t1);
1430 pair->push(t2);
1431 _packset.append(pair);
1432 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_use_defs: set_alignment(%d, %d, %d)", t1->_idx, t2->_idx, align);)
1433 set_alignment(t1, t2, align);
1434 changed = true;
1435 }
1436 }
1437 }
1438 return changed;
1439 }
1440
1441 //------------------------------follow_def_uses---------------------------
1442 // Extend the packset by visiting uses of nodes in pack p
follow_def_uses(Node_List * p)1443 bool SuperWord::follow_def_uses(Node_List* p) {
1444 bool changed = false;
1445 Node* s1 = p->at(0);
1446 Node* s2 = p->at(1);
1447 assert(p->size() == 2, "just checking");
1448 assert(s1->req() == s2->req(), "just checking");
1449 assert(alignment(s1) + data_size(s1) == alignment(s2), "just checking");
1450
1451 if (s1->is_Store()) return false;
1452
1453 int align = alignment(s1);
1454 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: s1 %d, align %d", s1->_idx, align);)
1455 int savings = -1;
1456 int num_s1_uses = 0;
1457 Node* u1 = NULL;
1458 Node* u2 = NULL;
1459 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1460 Node* t1 = s1->fast_out(i);
1461 num_s1_uses++;
1462 if (!in_bb(t1)) continue;
1463 for (DUIterator_Fast jmax, j = s2->fast_outs(jmax); j < jmax; j++) {
1464 Node* t2 = s2->fast_out(j);
1465 if (!in_bb(t2)) continue;
1466 if (t2->Opcode() == Op_AddI && t2 == _lp->as_CountedLoop()->incr()) continue; // don't mess with the iv
1467 if (!opnd_positions_match(s1, t1, s2, t2))
1468 continue;
1469 if (stmts_can_pack(t1, t2, align)) {
1470 int my_savings = est_savings(t1, t2);
1471 if (my_savings > savings) {
1472 savings = my_savings;
1473 u1 = t1;
1474 u2 = t2;
1475 }
1476 }
1477 }
1478 }
1479 if (num_s1_uses > 1) {
1480 _race_possible = true;
1481 }
1482 if (savings >= 0) {
1483 Node_List* pair = new Node_List();
1484 pair->push(u1);
1485 pair->push(u2);
1486 _packset.append(pair);
1487 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SuperWord::follow_def_uses: set_alignment(%d, %d, %d)", u1->_idx, u2->_idx, align);)
1488 set_alignment(u1, u2, align);
1489 changed = true;
1490 }
1491 return changed;
1492 }
1493
1494 //------------------------------order_def_uses---------------------------
1495 // For extended packsets, ordinally arrange uses packset by major component
order_def_uses(Node_List * p)1496 void SuperWord::order_def_uses(Node_List* p) {
1497 Node* s1 = p->at(0);
1498
1499 if (s1->is_Store()) return;
1500
1501 // reductions are always managed beforehand
1502 if (s1->is_reduction()) return;
1503
1504 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1505 Node* t1 = s1->fast_out(i);
1506
1507 // Only allow operand swap on commuting operations
1508 if (!t1->is_Add() && !t1->is_Mul() && !VectorNode::is_muladds2i(t1)) {
1509 break;
1510 }
1511
1512 // Now find t1's packset
1513 Node_List* p2 = NULL;
1514 for (int j = 0; j < _packset.length(); j++) {
1515 p2 = _packset.at(j);
1516 Node* first = p2->at(0);
1517 if (t1 == first) {
1518 break;
1519 }
1520 p2 = NULL;
1521 }
1522 // Arrange all sub components by the major component
1523 if (p2 != NULL) {
1524 for (uint j = 1; j < p->size(); j++) {
1525 Node* d1 = p->at(j);
1526 Node* u1 = p2->at(j);
1527 opnd_positions_match(s1, t1, d1, u1);
1528 }
1529 }
1530 }
1531 }
1532
1533 //---------------------------opnd_positions_match-------------------------
1534 // Is the use of d1 in u1 at the same operand position as d2 in u2?
opnd_positions_match(Node * d1,Node * u1,Node * d2,Node * u2)1535 bool SuperWord::opnd_positions_match(Node* d1, Node* u1, Node* d2, Node* u2) {
1536 // check reductions to see if they are marshalled to represent the reduction
1537 // operator in a specified opnd
1538 if (u1->is_reduction() && u2->is_reduction()) {
1539 // ensure reductions have phis and reduction definitions feeding the 1st operand
1540 Node* first = u1->in(2);
1541 if (first->is_Phi() || first->is_reduction()) {
1542 u1->swap_edges(1, 2);
1543 }
1544 // ensure reductions have phis and reduction definitions feeding the 1st operand
1545 first = u2->in(2);
1546 if (first->is_Phi() || first->is_reduction()) {
1547 u2->swap_edges(1, 2);
1548 }
1549 return true;
1550 }
1551
1552 uint ct = u1->req();
1553 if (ct != u2->req()) return false;
1554 uint i1 = 0;
1555 uint i2 = 0;
1556 do {
1557 for (i1++; i1 < ct; i1++) if (u1->in(i1) == d1) break;
1558 for (i2++; i2 < ct; i2++) if (u2->in(i2) == d2) break;
1559 if (i1 != i2) {
1560 if ((i1 == (3-i2)) && (u2->is_Add() || u2->is_Mul())) {
1561 // Further analysis relies on operands position matching.
1562 u2->swap_edges(i1, i2);
1563 } else if (VectorNode::is_muladds2i(u2) && u1 != u2) {
1564 if (i1 == 5 - i2) { // ((i1 == 3 && i2 == 2) || (i1 == 2 && i2 == 3) || (i1 == 1 && i2 == 4) || (i1 == 4 && i2 == 1))
1565 u2->swap_edges(1, 2);
1566 u2->swap_edges(3, 4);
1567 }
1568 if (i1 == 3 - i2 || i1 == 7 - i2) { // ((i1 == 1 && i2 == 2) || (i1 == 2 && i2 == 1) || (i1 == 3 && i2 == 4) || (i1 == 4 && i2 == 3))
1569 u2->swap_edges(2, 3);
1570 u2->swap_edges(1, 4);
1571 }
1572 return false; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
1573 } else {
1574 return false;
1575 }
1576 } else if (i1 == i2 && VectorNode::is_muladds2i(u2) && u1 != u2) {
1577 u2->swap_edges(1, 3);
1578 u2->swap_edges(2, 4);
1579 return false; // Just swap the edges, the muladds2i nodes get packed in follow_use_defs
1580 }
1581 } while (i1 < ct);
1582 return true;
1583 }
1584
1585 //------------------------------est_savings---------------------------
1586 // Estimate the savings from executing s1 and s2 as a pack
est_savings(Node * s1,Node * s2)1587 int SuperWord::est_savings(Node* s1, Node* s2) {
1588 int save_in = 2 - 1; // 2 operations per instruction in packed form
1589
1590 // inputs
1591 for (uint i = 1; i < s1->req(); i++) {
1592 Node* x1 = s1->in(i);
1593 Node* x2 = s2->in(i);
1594 if (x1 != x2) {
1595 if (are_adjacent_refs(x1, x2)) {
1596 save_in += adjacent_profit(x1, x2);
1597 } else if (!in_packset(x1, x2)) {
1598 save_in -= pack_cost(2);
1599 } else {
1600 save_in += unpack_cost(2);
1601 }
1602 }
1603 }
1604
1605 // uses of result
1606 uint ct = 0;
1607 int save_use = 0;
1608 for (DUIterator_Fast imax, i = s1->fast_outs(imax); i < imax; i++) {
1609 Node* s1_use = s1->fast_out(i);
1610 for (int j = 0; j < _packset.length(); j++) {
1611 Node_List* p = _packset.at(j);
1612 if (p->at(0) == s1_use) {
1613 for (DUIterator_Fast kmax, k = s2->fast_outs(kmax); k < kmax; k++) {
1614 Node* s2_use = s2->fast_out(k);
1615 if (p->at(p->size()-1) == s2_use) {
1616 ct++;
1617 if (are_adjacent_refs(s1_use, s2_use)) {
1618 save_use += adjacent_profit(s1_use, s2_use);
1619 }
1620 }
1621 }
1622 }
1623 }
1624 }
1625
1626 if (ct < s1->outcnt()) save_use += unpack_cost(1);
1627 if (ct < s2->outcnt()) save_use += unpack_cost(1);
1628
1629 return MAX2(save_in, save_use);
1630 }
1631
1632 //------------------------------costs---------------------------
adjacent_profit(Node * s1,Node * s2)1633 int SuperWord::adjacent_profit(Node* s1, Node* s2) { return 2; }
pack_cost(int ct)1634 int SuperWord::pack_cost(int ct) { return ct; }
unpack_cost(int ct)1635 int SuperWord::unpack_cost(int ct) { return ct; }
1636
1637 //------------------------------combine_packs---------------------------
1638 // Combine packs A and B with A.last == B.first into A.first..,A.last,B.second,..B.last
combine_packs()1639 void SuperWord::combine_packs() {
1640 bool changed = true;
1641 // Combine packs regardless max vector size.
1642 while (changed) {
1643 changed = false;
1644 for (int i = 0; i < _packset.length(); i++) {
1645 Node_List* p1 = _packset.at(i);
1646 if (p1 == NULL) continue;
1647 // Because of sorting we can start at i + 1
1648 for (int j = i + 1; j < _packset.length(); j++) {
1649 Node_List* p2 = _packset.at(j);
1650 if (p2 == NULL) continue;
1651 if (i == j) continue;
1652 if (p1->at(p1->size()-1) == p2->at(0)) {
1653 for (uint k = 1; k < p2->size(); k++) {
1654 p1->push(p2->at(k));
1655 }
1656 _packset.at_put(j, NULL);
1657 changed = true;
1658 }
1659 }
1660 }
1661 }
1662
1663 // Split packs which have size greater then max vector size.
1664 for (int i = 0; i < _packset.length(); i++) {
1665 Node_List* p1 = _packset.at(i);
1666 if (p1 != NULL) {
1667 BasicType bt = velt_basic_type(p1->at(0));
1668 uint max_vlen = Matcher::max_vector_size(bt); // Max elements in vector
1669 assert(is_power_of_2(max_vlen), "sanity");
1670 uint psize = p1->size();
1671 if (!is_power_of_2(psize)) {
1672 // Skip pack which can't be vector.
1673 // case1: for(...) { a[i] = i; } elements values are different (i+x)
1674 // case2: for(...) { a[i] = b[i+1]; } can't align both, load and store
1675 _packset.at_put(i, NULL);
1676 continue;
1677 }
1678 if (psize > max_vlen) {
1679 Node_List* pack = new Node_List();
1680 for (uint j = 0; j < psize; j++) {
1681 pack->push(p1->at(j));
1682 if (pack->size() >= max_vlen) {
1683 assert(is_power_of_2(pack->size()), "sanity");
1684 _packset.append(pack);
1685 pack = new Node_List();
1686 }
1687 }
1688 _packset.at_put(i, NULL);
1689 }
1690 }
1691 }
1692
1693 // Compress list.
1694 for (int i = _packset.length() - 1; i >= 0; i--) {
1695 Node_List* p1 = _packset.at(i);
1696 if (p1 == NULL) {
1697 _packset.remove_at(i);
1698 }
1699 }
1700
1701 if (TraceSuperWord) {
1702 tty->print_cr("\nAfter combine_packs");
1703 print_packset();
1704 }
1705 }
1706
1707 //-----------------------------construct_my_pack_map--------------------------
1708 // Construct the map from nodes to packs. Only valid after the
1709 // point where a node is only in one pack (after combine_packs).
construct_my_pack_map()1710 void SuperWord::construct_my_pack_map() {
1711 Node_List* rslt = NULL;
1712 for (int i = 0; i < _packset.length(); i++) {
1713 Node_List* p = _packset.at(i);
1714 for (uint j = 0; j < p->size(); j++) {
1715 Node* s = p->at(j);
1716 assert(my_pack(s) == NULL, "only in one pack");
1717 set_my_pack(s, p);
1718 }
1719 }
1720 }
1721
1722 //------------------------------filter_packs---------------------------
1723 // Remove packs that are not implemented or not profitable.
filter_packs()1724 void SuperWord::filter_packs() {
1725 // Remove packs that are not implemented
1726 for (int i = _packset.length() - 1; i >= 0; i--) {
1727 Node_List* pk = _packset.at(i);
1728 bool impl = implemented(pk);
1729 if (!impl) {
1730 #ifndef PRODUCT
1731 if (TraceSuperWord && Verbose) {
1732 tty->print_cr("Unimplemented");
1733 pk->at(0)->dump();
1734 }
1735 #endif
1736 remove_pack_at(i);
1737 }
1738 Node *n = pk->at(0);
1739 if (n->is_reduction()) {
1740 _num_reductions++;
1741 } else {
1742 _num_work_vecs++;
1743 }
1744 }
1745
1746 // Remove packs that are not profitable
1747 bool changed;
1748 do {
1749 changed = false;
1750 for (int i = _packset.length() - 1; i >= 0; i--) {
1751 Node_List* pk = _packset.at(i);
1752 bool prof = profitable(pk);
1753 if (!prof) {
1754 #ifndef PRODUCT
1755 if (TraceSuperWord && Verbose) {
1756 tty->print_cr("Unprofitable");
1757 pk->at(0)->dump();
1758 }
1759 #endif
1760 remove_pack_at(i);
1761 changed = true;
1762 }
1763 }
1764 } while (changed);
1765
1766 #ifndef PRODUCT
1767 if (TraceSuperWord) {
1768 tty->print_cr("\nAfter filter_packs");
1769 print_packset();
1770 tty->cr();
1771 }
1772 #endif
1773 }
1774
1775 //------------------------------merge_packs_to_cmovd---------------------------
1776 // Merge CMoveD into new vector-nodes
1777 // We want to catch this pattern and subsume CmpD and Bool into CMoveD
1778 //
1779 // SubD ConD
1780 // / | /
1781 // / | / /
1782 // / | / /
1783 // / | / /
1784 // / / /
1785 // / / | /
1786 // v / | /
1787 // CmpD | /
1788 // | | /
1789 // v | /
1790 // Bool | /
1791 // \ | /
1792 // \ | /
1793 // \ | /
1794 // \ | /
1795 // \ v /
1796 // CMoveD
1797 //
1798
merge_packs_to_cmovd()1799 void SuperWord::merge_packs_to_cmovd() {
1800 for (int i = _packset.length() - 1; i >= 0; i--) {
1801 _cmovev_kit.make_cmovevd_pack(_packset.at(i));
1802 }
1803 #ifndef PRODUCT
1804 if (TraceSuperWord) {
1805 tty->print_cr("\nSuperWord::merge_packs_to_cmovd(): After merge");
1806 print_packset();
1807 tty->cr();
1808 }
1809 #endif
1810 }
1811
is_Bool_candidate(Node * def) const1812 Node* CMoveKit::is_Bool_candidate(Node* def) const {
1813 Node* use = NULL;
1814 if (!def->is_Bool() || def->in(0) != NULL || def->outcnt() != 1) {
1815 return NULL;
1816 }
1817 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1818 use = def->fast_out(j);
1819 if (!_sw->same_generation(def, use) || !use->is_CMove()) {
1820 return NULL;
1821 }
1822 }
1823 return use;
1824 }
1825
is_CmpD_candidate(Node * def) const1826 Node* CMoveKit::is_CmpD_candidate(Node* def) const {
1827 Node* use = NULL;
1828 if (!def->is_Cmp() || def->in(0) != NULL || def->outcnt() != 1) {
1829 return NULL;
1830 }
1831 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1832 use = def->fast_out(j);
1833 if (!_sw->same_generation(def, use) || (use = is_Bool_candidate(use)) == NULL || !_sw->same_generation(def, use)) {
1834 return NULL;
1835 }
1836 }
1837 return use;
1838 }
1839
make_cmovevd_pack(Node_List * cmovd_pk)1840 Node_List* CMoveKit::make_cmovevd_pack(Node_List* cmovd_pk) {
1841 Node *cmovd = cmovd_pk->at(0);
1842 if (!cmovd->is_CMove()) {
1843 return NULL;
1844 }
1845 if (cmovd->Opcode() != Op_CMoveF && cmovd->Opcode() != Op_CMoveD) {
1846 return NULL;
1847 }
1848 if (pack(cmovd) != NULL) { // already in the cmov pack
1849 return NULL;
1850 }
1851 if (cmovd->in(0) != NULL) {
1852 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print("CMoveKit::make_cmovevd_pack: CMoveD %d has control flow, escaping...", cmovd->_idx); cmovd->dump();})
1853 return NULL;
1854 }
1855
1856 Node* bol = cmovd->as_CMove()->in(CMoveNode::Condition);
1857 if (!bol->is_Bool()
1858 || bol->outcnt() != 1
1859 || !_sw->same_generation(bol, cmovd)
1860 || bol->in(0) != NULL // BoolNode has control flow!!
1861 || _sw->my_pack(bol) == NULL) {
1862 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();})
1863 return NULL;
1864 }
1865 Node_List* bool_pk = _sw->my_pack(bol);
1866 if (bool_pk->size() != cmovd_pk->size() ) {
1867 return NULL;
1868 }
1869
1870 Node* cmpd = bol->in(1);
1871 if (!cmpd->is_Cmp()
1872 || cmpd->outcnt() != 1
1873 || !_sw->same_generation(cmpd, cmovd)
1874 || cmpd->in(0) != NULL // CmpDNode has control flow!!
1875 || _sw->my_pack(cmpd) == NULL) {
1876 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();})
1877 return NULL;
1878 }
1879 Node_List* cmpd_pk = _sw->my_pack(cmpd);
1880 if (cmpd_pk->size() != cmovd_pk->size() ) {
1881 return NULL;
1882 }
1883
1884 if (!test_cmpd_pack(cmpd_pk, cmovd_pk)) {
1885 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();})
1886 return NULL;
1887 }
1888
1889 Node_List* new_cmpd_pk = new Node_List();
1890 uint sz = cmovd_pk->size() - 1;
1891 for (uint i = 0; i <= sz; ++i) {
1892 Node* cmov = cmovd_pk->at(i);
1893 Node* bol = bool_pk->at(i);
1894 Node* cmp = cmpd_pk->at(i);
1895
1896 new_cmpd_pk->insert(i, cmov);
1897
1898 map(cmov, new_cmpd_pk);
1899 map(bol, new_cmpd_pk);
1900 map(cmp, new_cmpd_pk);
1901
1902 _sw->set_my_pack(cmov, new_cmpd_pk); // and keep old packs for cmp and bool
1903 }
1904 _sw->_packset.remove(cmovd_pk);
1905 _sw->_packset.remove(bool_pk);
1906 _sw->_packset.remove(cmpd_pk);
1907 _sw->_packset.append(new_cmpd_pk);
1908 NOT_PRODUCT(if(_sw->is_trace_cmov()) {tty->print_cr("CMoveKit::make_cmovevd_pack: added syntactic CMoveD pack"); _sw->print_pack(new_cmpd_pk);})
1909 return new_cmpd_pk;
1910 }
1911
test_cmpd_pack(Node_List * cmpd_pk,Node_List * cmovd_pk)1912 bool CMoveKit::test_cmpd_pack(Node_List* cmpd_pk, Node_List* cmovd_pk) {
1913 Node* cmpd0 = cmpd_pk->at(0);
1914 assert(cmpd0->is_Cmp(), "CMoveKit::test_cmpd_pack: should be CmpDNode");
1915 assert(cmovd_pk->at(0)->is_CMove(), "CMoveKit::test_cmpd_pack: should be CMoveD");
1916 assert(cmpd_pk->size() == cmovd_pk->size(), "CMoveKit::test_cmpd_pack: should be same size");
1917 Node* in1 = cmpd0->in(1);
1918 Node* in2 = cmpd0->in(2);
1919 Node_List* in1_pk = _sw->my_pack(in1);
1920 Node_List* in2_pk = _sw->my_pack(in2);
1921
1922 if ( (in1_pk != NULL && in1_pk->size() != cmpd_pk->size())
1923 || (in2_pk != NULL && in2_pk->size() != cmpd_pk->size()) ) {
1924 return false;
1925 }
1926
1927 // test if "all" in1 are in the same pack or the same node
1928 if (in1_pk == NULL) {
1929 for (uint j = 1; j < cmpd_pk->size(); j++) {
1930 if (cmpd_pk->at(j)->in(1) != in1) {
1931 return false;
1932 }
1933 }//for: in1_pk is not pack but all CmpD nodes in the pack have the same in(1)
1934 }
1935 // test if "all" in2 are in the same pack or the same node
1936 if (in2_pk == NULL) {
1937 for (uint j = 1; j < cmpd_pk->size(); j++) {
1938 if (cmpd_pk->at(j)->in(2) != in2) {
1939 return false;
1940 }
1941 }//for: in2_pk is not pack but all CmpD nodes in the pack have the same in(2)
1942 }
1943 //now check if cmpd_pk may be subsumed in vector built for cmovd_pk
1944 int cmovd_ind1, cmovd_ind2;
1945 if (cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1946 && cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1947 cmovd_ind1 = CMoveNode::IfFalse;
1948 cmovd_ind2 = CMoveNode::IfTrue;
1949 } else if (cmpd_pk->at(0)->in(2) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfFalse)
1950 && cmpd_pk->at(0)->in(1) == cmovd_pk->at(0)->as_CMove()->in(CMoveNode::IfTrue)) {
1951 cmovd_ind2 = CMoveNode::IfFalse;
1952 cmovd_ind1 = CMoveNode::IfTrue;
1953 }
1954 else {
1955 return false;
1956 }
1957
1958 for (uint j = 1; j < cmpd_pk->size(); j++) {
1959 if (cmpd_pk->at(j)->in(1) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind1)
1960 || cmpd_pk->at(j)->in(2) != cmovd_pk->at(j)->as_CMove()->in(cmovd_ind2)) {
1961 return false;
1962 }//if
1963 }
1964 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(); })
1965 return true;
1966 }
1967
1968 //------------------------------implemented---------------------------
1969 // Can code be generated for pack p?
implemented(Node_List * p)1970 bool SuperWord::implemented(Node_List* p) {
1971 bool retValue = false;
1972 Node* p0 = p->at(0);
1973 if (p0 != NULL) {
1974 int opc = p0->Opcode();
1975 uint size = p->size();
1976 if (p0->is_reduction()) {
1977 const Type *arith_type = p0->bottom_type();
1978 // Length 2 reductions of INT/LONG do not offer performance benefits
1979 if (((arith_type->basic_type() == T_INT) || (arith_type->basic_type() == T_LONG)) && (size == 2)) {
1980 retValue = false;
1981 } else {
1982 retValue = ReductionNode::implemented(opc, size, arith_type->basic_type());
1983 }
1984 } else {
1985 retValue = VectorNode::implemented(opc, size, velt_basic_type(p0));
1986 }
1987 if (!retValue) {
1988 if (is_cmov_pack(p)) {
1989 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::implemented: found cmpd pack"); print_pack(p);})
1990 return true;
1991 }
1992 }
1993 }
1994 return retValue;
1995 }
1996
is_cmov_pack(Node_List * p)1997 bool SuperWord::is_cmov_pack(Node_List* p) {
1998 return _cmovev_kit.pack(p->at(0)) != NULL;
1999 }
2000 //------------------------------same_inputs--------------------------
2001 // For pack p, are all idx operands the same?
same_inputs(Node_List * p,int idx)2002 bool SuperWord::same_inputs(Node_List* p, int idx) {
2003 Node* p0 = p->at(0);
2004 uint vlen = p->size();
2005 Node* p0_def = p0->in(idx);
2006 for (uint i = 1; i < vlen; i++) {
2007 Node* pi = p->at(i);
2008 Node* pi_def = pi->in(idx);
2009 if (p0_def != pi_def) {
2010 return false;
2011 }
2012 }
2013 return true;
2014 }
2015
2016 //------------------------------profitable---------------------------
2017 // For pack p, are all operands and all uses (with in the block) vector?
profitable(Node_List * p)2018 bool SuperWord::profitable(Node_List* p) {
2019 Node* p0 = p->at(0);
2020 uint start, end;
2021 VectorNode::vector_operands(p0, &start, &end);
2022
2023 // Return false if some inputs are not vectors or vectors with different
2024 // size or alignment.
2025 // Also, for now, return false if not scalar promotion case when inputs are
2026 // the same. Later, implement PackNode and allow differing, non-vector inputs
2027 // (maybe just the ones from outside the block.)
2028 for (uint i = start; i < end; i++) {
2029 if (!is_vector_use(p0, i)) {
2030 return false;
2031 }
2032 }
2033 // Check if reductions are connected
2034 if (p0->is_reduction()) {
2035 Node* second_in = p0->in(2);
2036 Node_List* second_pk = my_pack(second_in);
2037 if ((second_pk == NULL) || (_num_work_vecs == _num_reductions)) {
2038 // Remove reduction flag if no parent pack or if not enough work
2039 // to cover reduction expansion overhead
2040 p0->remove_flag(Node::Flag_is_reduction);
2041 return false;
2042 } else if (second_pk->size() != p->size()) {
2043 return false;
2044 }
2045 }
2046 if (VectorNode::is_shift(p0)) {
2047 // For now, return false if shift count is vector or not scalar promotion
2048 // case (different shift counts) because it is not supported yet.
2049 Node* cnt = p0->in(2);
2050 Node_List* cnt_pk = my_pack(cnt);
2051 if (cnt_pk != NULL)
2052 return false;
2053 if (!same_inputs(p, 2))
2054 return false;
2055 }
2056 if (!p0->is_Store()) {
2057 // For now, return false if not all uses are vector.
2058 // Later, implement ExtractNode and allow non-vector uses (maybe
2059 // just the ones outside the block.)
2060 for (uint i = 0; i < p->size(); i++) {
2061 Node* def = p->at(i);
2062 if (is_cmov_pack_internal_node(p, def)) {
2063 continue;
2064 }
2065 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2066 Node* use = def->fast_out(j);
2067 for (uint k = 0; k < use->req(); k++) {
2068 Node* n = use->in(k);
2069 if (def == n) {
2070 // Reductions should only have a Phi use at the loop head or a non-phi use
2071 // outside of the loop if it is the last element of the pack (e.g. SafePoint).
2072 if (def->is_reduction() &&
2073 ((use->is_Phi() && use->in(0) == _lpt->_head) ||
2074 (!_lpt->is_member(_phase->get_loop(_phase->ctrl_or_self(use))) && i == p->size()-1))) {
2075 continue;
2076 }
2077 if (!is_vector_use(use, k)) {
2078 return false;
2079 }
2080 }
2081 }
2082 }
2083 }
2084 }
2085 return true;
2086 }
2087
2088 //------------------------------schedule---------------------------
2089 // Adjust the memory graph for the packed operations
schedule()2090 void SuperWord::schedule() {
2091
2092 // Co-locate in the memory graph the members of each memory pack
2093 for (int i = 0; i < _packset.length(); i++) {
2094 co_locate_pack(_packset.at(i));
2095 }
2096 }
2097
2098 //-------------------------------remove_and_insert-------------------
2099 // Remove "current" from its current position in the memory graph and insert
2100 // it after the appropriate insertion point (lip or uip).
remove_and_insert(MemNode * current,MemNode * prev,MemNode * lip,Node * uip,Unique_Node_List & sched_before)2101 void SuperWord::remove_and_insert(MemNode *current, MemNode *prev, MemNode *lip,
2102 Node *uip, Unique_Node_List &sched_before) {
2103 Node* my_mem = current->in(MemNode::Memory);
2104 bool sched_up = sched_before.member(current);
2105
2106 // remove current_store from its current position in the memmory graph
2107 for (DUIterator i = current->outs(); current->has_out(i); i++) {
2108 Node* use = current->out(i);
2109 if (use->is_Mem()) {
2110 assert(use->in(MemNode::Memory) == current, "must be");
2111 if (use == prev) { // connect prev to my_mem
2112 _igvn.replace_input_of(use, MemNode::Memory, my_mem);
2113 --i; //deleted this edge; rescan position
2114 } else if (sched_before.member(use)) {
2115 if (!sched_up) { // Will be moved together with current
2116 _igvn.replace_input_of(use, MemNode::Memory, uip);
2117 --i; //deleted this edge; rescan position
2118 }
2119 } else {
2120 if (sched_up) { // Will be moved together with current
2121 _igvn.replace_input_of(use, MemNode::Memory, lip);
2122 --i; //deleted this edge; rescan position
2123 }
2124 }
2125 }
2126 }
2127
2128 Node *insert_pt = sched_up ? uip : lip;
2129
2130 // all uses of insert_pt's memory state should use current's instead
2131 for (DUIterator i = insert_pt->outs(); insert_pt->has_out(i); i++) {
2132 Node* use = insert_pt->out(i);
2133 if (use->is_Mem()) {
2134 assert(use->in(MemNode::Memory) == insert_pt, "must be");
2135 _igvn.replace_input_of(use, MemNode::Memory, current);
2136 --i; //deleted this edge; rescan position
2137 } else if (!sched_up && use->is_Phi() && use->bottom_type() == Type::MEMORY) {
2138 uint pos; //lip (lower insert point) must be the last one in the memory slice
2139 for (pos=1; pos < use->req(); pos++) {
2140 if (use->in(pos) == insert_pt) break;
2141 }
2142 _igvn.replace_input_of(use, pos, current);
2143 --i;
2144 }
2145 }
2146
2147 //connect current to insert_pt
2148 _igvn.replace_input_of(current, MemNode::Memory, insert_pt);
2149 }
2150
2151 //------------------------------co_locate_pack----------------------------------
2152 // To schedule a store pack, we need to move any sandwiched memory ops either before
2153 // or after the pack, based upon dependence information:
2154 // (1) If any store in the pack depends on the sandwiched memory op, the
2155 // sandwiched memory op must be scheduled BEFORE the pack;
2156 // (2) If a sandwiched memory op depends on any store in the pack, the
2157 // sandwiched memory op must be scheduled AFTER the pack;
2158 // (3) If a sandwiched memory op (say, memA) depends on another sandwiched
2159 // memory op (say memB), memB must be scheduled before memA. So, if memA is
2160 // scheduled before the pack, memB must also be scheduled before the pack;
2161 // (4) If there is no dependence restriction for a sandwiched memory op, we simply
2162 // schedule this store AFTER the pack
2163 // (5) We know there is no dependence cycle, so there in no other case;
2164 // (6) Finally, all memory ops in another single pack should be moved in the same direction.
2165 //
2166 // To schedule a load pack, we use the memory state of either the first or the last load in
2167 // the pack, based on the dependence constraint.
co_locate_pack(Node_List * pk)2168 void SuperWord::co_locate_pack(Node_List* pk) {
2169 if (pk->at(0)->is_Store()) {
2170 MemNode* first = executed_first(pk)->as_Mem();
2171 MemNode* last = executed_last(pk)->as_Mem();
2172 Unique_Node_List schedule_before_pack;
2173 Unique_Node_List memops;
2174
2175 MemNode* current = last->in(MemNode::Memory)->as_Mem();
2176 MemNode* previous = last;
2177 while (true) {
2178 assert(in_bb(current), "stay in block");
2179 memops.push(previous);
2180 for (DUIterator i = current->outs(); current->has_out(i); i++) {
2181 Node* use = current->out(i);
2182 if (use->is_Mem() && use != previous)
2183 memops.push(use);
2184 }
2185 if (current == first) break;
2186 previous = current;
2187 current = current->in(MemNode::Memory)->as_Mem();
2188 }
2189
2190 // determine which memory operations should be scheduled before the pack
2191 for (uint i = 1; i < memops.size(); i++) {
2192 Node *s1 = memops.at(i);
2193 if (!in_pack(s1, pk) && !schedule_before_pack.member(s1)) {
2194 for (uint j = 0; j< i; j++) {
2195 Node *s2 = memops.at(j);
2196 if (!independent(s1, s2)) {
2197 if (in_pack(s2, pk) || schedule_before_pack.member(s2)) {
2198 schedule_before_pack.push(s1); // s1 must be scheduled before
2199 Node_List* mem_pk = my_pack(s1);
2200 if (mem_pk != NULL) {
2201 for (uint ii = 0; ii < mem_pk->size(); ii++) {
2202 Node* s = mem_pk->at(ii); // follow partner
2203 if (memops.member(s) && !schedule_before_pack.member(s))
2204 schedule_before_pack.push(s);
2205 }
2206 }
2207 break;
2208 }
2209 }
2210 }
2211 }
2212 }
2213
2214 Node* upper_insert_pt = first->in(MemNode::Memory);
2215 // Following code moves loads connected to upper_insert_pt below aliased stores.
2216 // Collect such loads here and reconnect them back to upper_insert_pt later.
2217 memops.clear();
2218 for (DUIterator i = upper_insert_pt->outs(); upper_insert_pt->has_out(i); i++) {
2219 Node* use = upper_insert_pt->out(i);
2220 if (use->is_Mem() && !use->is_Store()) {
2221 memops.push(use);
2222 }
2223 }
2224
2225 MemNode* lower_insert_pt = last;
2226 previous = last; //previous store in pk
2227 current = last->in(MemNode::Memory)->as_Mem();
2228
2229 // start scheduling from "last" to "first"
2230 while (true) {
2231 assert(in_bb(current), "stay in block");
2232 assert(in_pack(previous, pk), "previous stays in pack");
2233 Node* my_mem = current->in(MemNode::Memory);
2234
2235 if (in_pack(current, pk)) {
2236 // Forward users of my memory state (except "previous) to my input memory state
2237 for (DUIterator i = current->outs(); current->has_out(i); i++) {
2238 Node* use = current->out(i);
2239 if (use->is_Mem() && use != previous) {
2240 assert(use->in(MemNode::Memory) == current, "must be");
2241 if (schedule_before_pack.member(use)) {
2242 _igvn.replace_input_of(use, MemNode::Memory, upper_insert_pt);
2243 } else {
2244 _igvn.replace_input_of(use, MemNode::Memory, lower_insert_pt);
2245 }
2246 --i; // deleted this edge; rescan position
2247 }
2248 }
2249 previous = current;
2250 } else { // !in_pack(current, pk) ==> a sandwiched store
2251 remove_and_insert(current, previous, lower_insert_pt, upper_insert_pt, schedule_before_pack);
2252 }
2253
2254 if (current == first) break;
2255 current = my_mem->as_Mem();
2256 } // end while
2257
2258 // Reconnect loads back to upper_insert_pt.
2259 for (uint i = 0; i < memops.size(); i++) {
2260 Node *ld = memops.at(i);
2261 if (ld->in(MemNode::Memory) != upper_insert_pt) {
2262 _igvn.replace_input_of(ld, MemNode::Memory, upper_insert_pt);
2263 }
2264 }
2265 } else if (pk->at(0)->is_Load()) { // Load pack
2266 // All loads in the pack should have the same memory state. By default,
2267 // we use the memory state of the last load. However, if any load could
2268 // not be moved down due to the dependence constraint, we use the memory
2269 // state of the first load.
2270 Node* mem_input = pick_mem_state(pk);
2271 _igvn.hash_delete(mem_input);
2272 // Give each load the same memory state
2273 for (uint i = 0; i < pk->size(); i++) {
2274 LoadNode* ld = pk->at(i)->as_Load();
2275 _igvn.replace_input_of(ld, MemNode::Memory, mem_input);
2276 }
2277 }
2278 }
2279
2280 // Finds the first and last memory state and then picks either of them by checking dependence constraints.
2281 // If a store is dependent on an earlier load then we need to pick the memory state of the first load and cannot
2282 // pick the memory state of the last load.
pick_mem_state(Node_List * pk)2283 Node* SuperWord::pick_mem_state(Node_List* pk) {
2284 Node* first_mem = find_first_mem_state(pk);
2285 Node* last_mem = find_last_mem_state(pk, first_mem);
2286
2287 for (uint i = 0; i < pk->size(); i++) {
2288 Node* ld = pk->at(i);
2289 for (Node* current = last_mem; current != ld->in(MemNode::Memory); current = current->in(MemNode::Memory)) {
2290 assert(current->is_Mem() && in_bb(current), "unexpected memory");
2291 assert(current != first_mem, "corrupted memory graph");
2292 if (!independent(current, ld)) {
2293 #ifdef ASSERT
2294 // Added assertion code since no case has been observed that should pick the first memory state.
2295 // Remove the assertion code whenever we find a (valid) case that really needs the first memory state.
2296 pk->dump();
2297 first_mem->dump();
2298 last_mem->dump();
2299 current->dump();
2300 ld->dump();
2301 ld->in(MemNode::Memory)->dump();
2302 assert(false, "never observed that first memory should be picked");
2303 #endif
2304 return first_mem; // A later store depends on this load, pick memory state of first load
2305 }
2306 }
2307 }
2308 return last_mem;
2309 }
2310
2311 // Walk the memory graph from the current first load until the
2312 // start of the loop and check if nodes on the way are memory
2313 // edges of loads in the pack. The last one we encounter is the
2314 // first load.
find_first_mem_state(Node_List * pk)2315 Node* SuperWord::find_first_mem_state(Node_List* pk) {
2316 Node* first_mem = pk->at(0)->in(MemNode::Memory);
2317 for (Node* current = first_mem; in_bb(current); current = current->is_Phi() ? current->in(LoopNode::EntryControl) : current->in(MemNode::Memory)) {
2318 assert(current->is_Mem() || (current->is_Phi() && current->in(0) == bb()), "unexpected memory");
2319 for (uint i = 1; i < pk->size(); i++) {
2320 Node* ld = pk->at(i);
2321 if (ld->in(MemNode::Memory) == current) {
2322 first_mem = current;
2323 break;
2324 }
2325 }
2326 }
2327 return first_mem;
2328 }
2329
2330 // Find the last load by going over the pack again and walking
2331 // the memory graph from the loads of the pack to the memory of
2332 // the first load. If we encounter the memory of the current last
2333 // load, then we started from further down in the memory graph and
2334 // the load we started from is the last load.
find_last_mem_state(Node_List * pk,Node * first_mem)2335 Node* SuperWord::find_last_mem_state(Node_List* pk, Node* first_mem) {
2336 Node* last_mem = pk->at(0)->in(MemNode::Memory);
2337 for (uint i = 0; i < pk->size(); i++) {
2338 Node* ld = pk->at(i);
2339 for (Node* current = ld->in(MemNode::Memory); current != first_mem; current = current->in(MemNode::Memory)) {
2340 assert(current->is_Mem() && in_bb(current), "unexpected memory");
2341 if (current->in(MemNode::Memory) == last_mem) {
2342 last_mem = ld->in(MemNode::Memory);
2343 }
2344 }
2345 }
2346 return last_mem;
2347 }
2348
2349 #ifndef PRODUCT
print_loop(bool whole)2350 void SuperWord::print_loop(bool whole) {
2351 Node_Stack stack(_arena, _phase->C->unique() >> 2);
2352 Node_List rpo_list;
2353 VectorSet visited(_arena);
2354 visited.set(lpt()->_head->_idx);
2355 _phase->rpo(lpt()->_head, stack, visited, rpo_list);
2356 _phase->dump(lpt(), rpo_list.size(), rpo_list );
2357 if(whole) {
2358 tty->print_cr("\n Whole loop tree");
2359 _phase->dump();
2360 tty->print_cr(" End of whole loop tree\n");
2361 }
2362 }
2363 #endif
2364
2365 //------------------------------output---------------------------
2366 // Convert packs into vector node operations
output()2367 void SuperWord::output() {
2368 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2369 Compile* C = _phase->C;
2370 if (_packset.length() == 0) {
2371 if (cl->is_main_loop()) {
2372 // Instigate more unrolling for optimization when vectorization fails.
2373 C->set_major_progress();
2374 cl->set_notpassed_slp();
2375 cl->mark_do_unroll_only();
2376 }
2377 return;
2378 }
2379
2380 #ifndef PRODUCT
2381 if (TraceLoopOpts) {
2382 tty->print("SuperWord::output ");
2383 lpt()->dump_head();
2384 }
2385 #endif
2386
2387 if (cl->is_main_loop()) {
2388 // MUST ENSURE main loop's initial value is properly aligned:
2389 // (iv_initial_value + min_iv_offset) % vector_width_in_bytes() == 0
2390
2391 align_initial_loop_index(align_to_ref());
2392
2393 // Insert extract (unpack) operations for scalar uses
2394 for (int i = 0; i < _packset.length(); i++) {
2395 insert_extracts(_packset.at(i));
2396 }
2397 }
2398
2399 uint max_vlen_in_bytes = 0;
2400 uint max_vlen = 0;
2401 bool can_process_post_loop = (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop());
2402
2403 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop before create_reserve_version_of_loop"); print_loop(true);})
2404
2405 CountedLoopReserveKit make_reversable(_phase, _lpt, do_reserve_copy());
2406
2407 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("SWPointer::output: print loop after create_reserve_version_of_loop"); print_loop(true);})
2408
2409 if (do_reserve_copy() && !make_reversable.has_reserved()) {
2410 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: loop was not reserved correctly, exiting SuperWord");})
2411 return;
2412 }
2413
2414 for (int i = 0; i < _block.length(); i++) {
2415 Node* n = _block.at(i);
2416 Node_List* p = my_pack(n);
2417 if (p && n == executed_last(p)) {
2418 uint vlen = p->size();
2419 uint vlen_in_bytes = 0;
2420 Node* vn = NULL;
2421 Node* low_adr = p->at(0);
2422 Node* first = executed_first(p);
2423 if (can_process_post_loop) {
2424 // override vlen with the main loops vector length
2425 vlen = cl->slp_max_unroll();
2426 }
2427 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);})
2428 int opc = n->Opcode();
2429 if (n->is_Load()) {
2430 Node* ctl = n->in(MemNode::Control);
2431 Node* mem = first->in(MemNode::Memory);
2432 SWPointer p1(n->as_Mem(), this, NULL, false);
2433 // Identify the memory dependency for the new loadVector node by
2434 // walking up through memory chain.
2435 // This is done to give flexibility to the new loadVector node so that
2436 // it can move above independent storeVector nodes.
2437 while (mem->is_StoreVector()) {
2438 SWPointer p2(mem->as_Mem(), this, NULL, false);
2439 int cmp = p1.cmp(p2);
2440 if (SWPointer::not_equal(cmp) || !SWPointer::comparable(cmp)) {
2441 mem = mem->in(MemNode::Memory);
2442 } else {
2443 break; // dependent memory
2444 }
2445 }
2446 Node* adr = low_adr->in(MemNode::Address);
2447 const TypePtr* atyp = n->adr_type();
2448 vn = LoadVectorNode::make(opc, ctl, mem, adr, atyp, vlen, velt_basic_type(n), control_dependency(p));
2449 vlen_in_bytes = vn->as_LoadVector()->memory_size();
2450 } else if (n->is_Store()) {
2451 // Promote value to be stored to vector
2452 Node* val = vector_opd(p, MemNode::ValueIn);
2453 if (val == NULL) {
2454 if (do_reserve_copy()) {
2455 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: val should not be NULL, exiting SuperWord");})
2456 return; //and reverse to backup IG
2457 }
2458 ShouldNotReachHere();
2459 }
2460
2461 Node* ctl = n->in(MemNode::Control);
2462 Node* mem = first->in(MemNode::Memory);
2463 Node* adr = low_adr->in(MemNode::Address);
2464 const TypePtr* atyp = n->adr_type();
2465 vn = StoreVectorNode::make(opc, ctl, mem, adr, atyp, val, vlen);
2466 vlen_in_bytes = vn->as_StoreVector()->memory_size();
2467 } else if (VectorNode::is_roundopD(n)) {
2468 Node* in1 = vector_opd(p, 1);
2469 Node* in2 = low_adr->in(2);
2470 assert(in2->is_Con(), "Constant rounding mode expected.");
2471 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2472 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2473 } else if (VectorNode::is_muladds2i(n)) {
2474 assert(n->req() == 5u, "MulAddS2I should have 4 operands.");
2475 Node* in1 = vector_opd(p, 1);
2476 Node* in2 = vector_opd(p, 2);
2477 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2478 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2479 } else if (n->req() == 3 && !is_cmov_pack(p)) {
2480 // Promote operands to vector
2481 Node* in1 = NULL;
2482 bool node_isa_reduction = n->is_reduction();
2483 if (node_isa_reduction) {
2484 // the input to the first reduction operation is retained
2485 in1 = low_adr->in(1);
2486 } else {
2487 in1 = vector_opd(p, 1);
2488 if (in1 == NULL) {
2489 if (do_reserve_copy()) {
2490 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in1 should not be NULL, exiting SuperWord");})
2491 return; //and reverse to backup IG
2492 }
2493 ShouldNotReachHere();
2494 }
2495 }
2496 Node* in2 = vector_opd(p, 2);
2497 if (in2 == NULL) {
2498 if (do_reserve_copy()) {
2499 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: in2 should not be NULL, exiting SuperWord");})
2500 return; //and reverse to backup IG
2501 }
2502 ShouldNotReachHere();
2503 }
2504 if (VectorNode::is_invariant_vector(in1) && (node_isa_reduction == false) && (n->is_Add() || n->is_Mul())) {
2505 // Move invariant vector input into second position to avoid register spilling.
2506 Node* tmp = in1;
2507 in1 = in2;
2508 in2 = tmp;
2509 }
2510 if (node_isa_reduction) {
2511 const Type *arith_type = n->bottom_type();
2512 vn = ReductionNode::make(opc, NULL, in1, in2, arith_type->basic_type());
2513 if (in2->is_Load()) {
2514 vlen_in_bytes = in2->as_LoadVector()->memory_size();
2515 } else {
2516 vlen_in_bytes = in2->as_Vector()->length_in_bytes();
2517 }
2518 } else {
2519 vn = VectorNode::make(opc, in1, in2, vlen, velt_basic_type(n));
2520 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2521 }
2522 } else if (opc == Op_SqrtF || opc == Op_SqrtD ||
2523 opc == Op_AbsF || opc == Op_AbsD ||
2524 opc == Op_AbsI || opc == Op_AbsL ||
2525 opc == Op_NegF || opc == Op_NegD ||
2526 opc == Op_PopCountI) {
2527 assert(n->req() == 2, "only one input expected");
2528 Node* in = vector_opd(p, 1);
2529 vn = VectorNode::make(opc, in, NULL, vlen, velt_basic_type(n));
2530 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2531 } else if (is_cmov_pack(p)) {
2532 if (can_process_post_loop) {
2533 // do not refactor of flow in post loop context
2534 return;
2535 }
2536 if (!n->is_CMove()) {
2537 continue;
2538 }
2539 // place here CMoveVDNode
2540 NOT_PRODUCT(if(is_trace_cmov()) {tty->print_cr("SWPointer::output: print before CMove vectorization"); print_loop(false);})
2541 Node* bol = n->in(CMoveNode::Condition);
2542 if (!bol->is_Bool() && bol->Opcode() == Op_ExtractI && bol->req() > 1 ) {
2543 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();})
2544 bol = bol->in(1); //may be ExtractNode
2545 }
2546
2547 assert(bol->is_Bool(), "should be BoolNode - too late to bail out!");
2548 if (!bol->is_Bool()) {
2549 if (do_reserve_copy()) {
2550 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: expected %d bool node, exiting SuperWord", bol->_idx); bol->dump();})
2551 return; //and reverse to backup IG
2552 }
2553 ShouldNotReachHere();
2554 }
2555
2556 int cond = (int)bol->as_Bool()->_test._test;
2557 Node* in_cc = _igvn.intcon(cond);
2558 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created intcon in_cc node %d", in_cc->_idx); in_cc->dump();})
2559 Node* cc = bol->clone();
2560 cc->set_req(1, in_cc);
2561 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created bool cc node %d", cc->_idx); cc->dump();})
2562
2563 Node* src1 = vector_opd(p, 2); //2=CMoveNode::IfFalse
2564 if (src1 == NULL) {
2565 if (do_reserve_copy()) {
2566 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src1 should not be NULL, exiting SuperWord");})
2567 return; //and reverse to backup IG
2568 }
2569 ShouldNotReachHere();
2570 }
2571 Node* src2 = vector_opd(p, 3); //3=CMoveNode::IfTrue
2572 if (src2 == NULL) {
2573 if (do_reserve_copy()) {
2574 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: src2 should not be NULL, exiting SuperWord");})
2575 return; //and reverse to backup IG
2576 }
2577 ShouldNotReachHere();
2578 }
2579 BasicType bt = velt_basic_type(n);
2580 const TypeVect* vt = TypeVect::make(bt, vlen);
2581 assert(bt == T_FLOAT || bt == T_DOUBLE, "Only vectorization for FP cmovs is supported");
2582 if (bt == T_FLOAT) {
2583 vn = new CMoveVFNode(cc, src1, src2, vt);
2584 } else {
2585 assert(bt == T_DOUBLE, "Expected double");
2586 vn = new CMoveVDNode(cc, src1, src2, vt);
2587 }
2588 NOT_PRODUCT(if(is_trace_cmov()) {tty->print("SWPointer::output: created new CMove node %d: ", vn->_idx); vn->dump();})
2589 } else if (opc == Op_FmaD || opc == Op_FmaF) {
2590 // Promote operands to vector
2591 Node* in1 = vector_opd(p, 1);
2592 Node* in2 = vector_opd(p, 2);
2593 Node* in3 = vector_opd(p, 3);
2594 vn = VectorNode::make(opc, in1, in2, in3, vlen, velt_basic_type(n));
2595 vlen_in_bytes = vn->as_Vector()->length_in_bytes();
2596 } else {
2597 if (do_reserve_copy()) {
2598 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: ShouldNotReachHere, exiting SuperWord");})
2599 return; //and reverse to backup IG
2600 }
2601 ShouldNotReachHere();
2602 }
2603
2604 assert(vn != NULL, "sanity");
2605 if (vn == NULL) {
2606 if (do_reserve_copy()){
2607 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("SWPointer::output: got NULL node, cannot proceed, exiting SuperWord");})
2608 return; //and reverse to backup IG
2609 }
2610 ShouldNotReachHere();
2611 }
2612
2613 _block.at_put(i, vn);
2614 _igvn.register_new_node_with_optimizer(vn);
2615 _phase->set_ctrl(vn, _phase->get_ctrl(p->at(0)));
2616 for (uint j = 0; j < p->size(); j++) {
2617 Node* pm = p->at(j);
2618 _igvn.replace_node(pm, vn);
2619 }
2620 _igvn._worklist.push(vn);
2621
2622 if (can_process_post_loop) {
2623 // first check if the vector size if the maximum vector which we can use on the machine,
2624 // other vector size have reduced values for predicated data mapping.
2625 if (vlen_in_bytes != (uint)MaxVectorSize) {
2626 return;
2627 }
2628 }
2629
2630 if (vlen > max_vlen) {
2631 max_vlen = vlen;
2632 }
2633 if (vlen_in_bytes > max_vlen_in_bytes) {
2634 max_vlen_in_bytes = vlen_in_bytes;
2635 }
2636 #ifdef ASSERT
2637 if (TraceNewVectors) {
2638 tty->print("new Vector node: ");
2639 vn->dump();
2640 }
2641 #endif
2642 }
2643 }//for (int i = 0; i < _block.length(); i++)
2644
2645 if (max_vlen_in_bytes > C->max_vector_size()) {
2646 C->set_max_vector_size(max_vlen_in_bytes);
2647 }
2648 if (max_vlen_in_bytes > 0) {
2649 cl->mark_loop_vectorized();
2650 }
2651
2652 if (SuperWordLoopUnrollAnalysis) {
2653 if (cl->has_passed_slp()) {
2654 uint slp_max_unroll_factor = cl->slp_max_unroll();
2655 if (slp_max_unroll_factor == max_vlen) {
2656 if (TraceSuperWordLoopUnrollAnalysis) {
2657 tty->print_cr("vector loop(unroll=%d, len=%d)\n", max_vlen, max_vlen_in_bytes*BitsPerByte);
2658 }
2659
2660 // For atomic unrolled loops which are vector mapped, instigate more unrolling
2661 cl->set_notpassed_slp();
2662 if (cl->is_main_loop()) {
2663 // if vector resources are limited, do not allow additional unrolling, also
2664 // do not unroll more on pure vector loops which were not reduced so that we can
2665 // program the post loop to single iteration execution.
2666 if (FLOATPRESSURE > 8) {
2667 C->set_major_progress();
2668 cl->mark_do_unroll_only();
2669 }
2670 }
2671
2672 if (do_reserve_copy()) {
2673 if (can_process_post_loop) {
2674 // Now create the difference of trip and limit and use it as our mask index.
2675 // Note: We limited the unroll of the vectorized loop so that
2676 // only vlen-1 size iterations can remain to be mask programmed.
2677 Node *incr = cl->incr();
2678 SubINode *index = new SubINode(cl->limit(), cl->init_trip());
2679 _igvn.register_new_node_with_optimizer(index);
2680 SetVectMaskINode *mask = new SetVectMaskINode(_phase->get_ctrl(cl->init_trip()), index);
2681 _igvn.register_new_node_with_optimizer(mask);
2682 // make this a single iteration loop
2683 AddINode *new_incr = new AddINode(incr->in(1), mask);
2684 _igvn.register_new_node_with_optimizer(new_incr);
2685 _phase->set_ctrl(new_incr, _phase->get_ctrl(incr));
2686 _igvn.replace_node(incr, new_incr);
2687 cl->mark_is_multiversioned();
2688 cl->loopexit()->add_flag(Node::Flag_has_vector_mask_set);
2689 }
2690 }
2691 }
2692 }
2693 }
2694
2695 if (do_reserve_copy()) {
2696 make_reversable.use_new();
2697 }
2698 NOT_PRODUCT(if(is_trace_loop_reverse()) {tty->print_cr("\n Final loop after SuperWord"); print_loop(true);})
2699 return;
2700 }
2701
2702 //------------------------------vector_opd---------------------------
2703 // Create a vector operand for the nodes in pack p for operand: in(opd_idx)
vector_opd(Node_List * p,int opd_idx)2704 Node* SuperWord::vector_opd(Node_List* p, int opd_idx) {
2705 Node* p0 = p->at(0);
2706 uint vlen = p->size();
2707 Node* opd = p0->in(opd_idx);
2708 CountedLoopNode *cl = lpt()->_head->as_CountedLoop();
2709
2710 if (PostLoopMultiversioning && Matcher::has_predicated_vectors() && cl->is_post_loop()) {
2711 // override vlen with the main loops vector length
2712 vlen = cl->slp_max_unroll();
2713 }
2714
2715 if (same_inputs(p, opd_idx)) {
2716 if (opd->is_Vector() || opd->is_LoadVector()) {
2717 assert(((opd_idx != 2) || !VectorNode::is_shift(p0)), "shift's count can't be vector");
2718 if (opd_idx == 2 && VectorNode::is_shift(p0)) {
2719 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("shift's count can't be vector");})
2720 return NULL;
2721 }
2722 return opd; // input is matching vector
2723 }
2724 if ((opd_idx == 2) && VectorNode::is_shift(p0)) {
2725 Compile* C = _phase->C;
2726 Node* cnt = opd;
2727 // Vector instructions do not mask shift count, do it here.
2728 juint mask = (p0->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2729 const TypeInt* t = opd->find_int_type();
2730 if (t != NULL && t->is_con()) {
2731 juint shift = t->get_con();
2732 if (shift > mask) { // Unsigned cmp
2733 cnt = ConNode::make(TypeInt::make(shift & mask));
2734 }
2735 } else {
2736 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2737 cnt = ConNode::make(TypeInt::make(mask));
2738 _igvn.register_new_node_with_optimizer(cnt);
2739 cnt = new AndINode(opd, cnt);
2740 _igvn.register_new_node_with_optimizer(cnt);
2741 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2742 }
2743 assert(opd->bottom_type()->isa_int(), "int type only");
2744 if (!opd->bottom_type()->isa_int()) {
2745 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should be int type only");})
2746 return NULL;
2747 }
2748 }
2749 // Move shift count into vector register.
2750 cnt = VectorNode::shift_count(p0, cnt, vlen, velt_basic_type(p0));
2751 _igvn.register_new_node_with_optimizer(cnt);
2752 _phase->set_ctrl(cnt, _phase->get_ctrl(opd));
2753 return cnt;
2754 }
2755 assert(!opd->is_StoreVector(), "such vector is not expected here");
2756 if (opd->is_StoreVector()) {
2757 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("StoreVector is not expected here");})
2758 return NULL;
2759 }
2760 // Convert scalar input to vector with the same number of elements as
2761 // p0's vector. Use p0's type because size of operand's container in
2762 // vector should match p0's size regardless operand's size.
2763 const Type* p0_t = velt_type(p0);
2764 VectorNode* vn = VectorNode::scalar2vector(opd, vlen, p0_t);
2765
2766 _igvn.register_new_node_with_optimizer(vn);
2767 _phase->set_ctrl(vn, _phase->get_ctrl(opd));
2768 #ifdef ASSERT
2769 if (TraceNewVectors) {
2770 tty->print("new Vector node: ");
2771 vn->dump();
2772 }
2773 #endif
2774 return vn;
2775 }
2776
2777 // Insert pack operation
2778 BasicType bt = velt_basic_type(p0);
2779 PackNode* pk = PackNode::make(opd, vlen, bt);
2780 DEBUG_ONLY( const BasicType opd_bt = opd->bottom_type()->basic_type(); )
2781
2782 for (uint i = 1; i < vlen; i++) {
2783 Node* pi = p->at(i);
2784 Node* in = pi->in(opd_idx);
2785 assert(my_pack(in) == NULL, "Should already have been unpacked");
2786 if (my_pack(in) != NULL) {
2787 NOT_PRODUCT(if(is_trace_loop_reverse() || TraceLoopOpts) {tty->print_cr("Should already have been unpacked");})
2788 return NULL;
2789 }
2790 assert(opd_bt == in->bottom_type()->basic_type(), "all same type");
2791 pk->add_opd(in);
2792 if (VectorNode::is_muladds2i(pi)) {
2793 Node* in2 = pi->in(opd_idx + 2);
2794 assert(my_pack(in2) == NULL, "Should already have been unpacked");
2795 if (my_pack(in2) != NULL) {
2796 NOT_PRODUCT(if (is_trace_loop_reverse() || TraceLoopOpts) { tty->print_cr("Should already have been unpacked"); })
2797 return NULL;
2798 }
2799 assert(opd_bt == in2->bottom_type()->basic_type(), "all same type");
2800 pk->add_opd(in2);
2801 }
2802 }
2803 _igvn.register_new_node_with_optimizer(pk);
2804 _phase->set_ctrl(pk, _phase->get_ctrl(opd));
2805 #ifdef ASSERT
2806 if (TraceNewVectors) {
2807 tty->print("new Vector node: ");
2808 pk->dump();
2809 }
2810 #endif
2811 return pk;
2812 }
2813
2814 //------------------------------insert_extracts---------------------------
2815 // If a use of pack p is not a vector use, then replace the
2816 // use with an extract operation.
insert_extracts(Node_List * p)2817 void SuperWord::insert_extracts(Node_List* p) {
2818 if (p->at(0)->is_Store()) return;
2819 assert(_n_idx_list.is_empty(), "empty (node,index) list");
2820
2821 // Inspect each use of each pack member. For each use that is
2822 // not a vector use, replace the use with an extract operation.
2823
2824 for (uint i = 0; i < p->size(); i++) {
2825 Node* def = p->at(i);
2826 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2827 Node* use = def->fast_out(j);
2828 for (uint k = 0; k < use->req(); k++) {
2829 Node* n = use->in(k);
2830 if (def == n) {
2831 Node_List* u_pk = my_pack(use);
2832 if ((u_pk == NULL || !is_cmov_pack(u_pk) || use->is_CMove()) && !is_vector_use(use, k)) {
2833 _n_idx_list.push(use, k);
2834 }
2835 }
2836 }
2837 }
2838 }
2839
2840 while (_n_idx_list.is_nonempty()) {
2841 Node* use = _n_idx_list.node();
2842 int idx = _n_idx_list.index();
2843 _n_idx_list.pop();
2844 Node* def = use->in(idx);
2845
2846 if (def->is_reduction()) continue;
2847
2848 // Insert extract operation
2849 _igvn.hash_delete(def);
2850 int def_pos = alignment(def) / data_size(def);
2851
2852 Node* ex = ExtractNode::make(def, def_pos, velt_basic_type(def));
2853 _igvn.register_new_node_with_optimizer(ex);
2854 _phase->set_ctrl(ex, _phase->get_ctrl(def));
2855 _igvn.replace_input_of(use, idx, ex);
2856 _igvn._worklist.push(def);
2857
2858 bb_insert_after(ex, bb_idx(def));
2859 set_velt_type(ex, velt_type(def));
2860 }
2861 }
2862
2863 //------------------------------is_vector_use---------------------------
2864 // Is use->in(u_idx) a vector use?
is_vector_use(Node * use,int u_idx)2865 bool SuperWord::is_vector_use(Node* use, int u_idx) {
2866 Node_List* u_pk = my_pack(use);
2867 if (u_pk == NULL) return false;
2868 if (use->is_reduction()) return true;
2869 Node* def = use->in(u_idx);
2870 Node_List* d_pk = my_pack(def);
2871 if (d_pk == NULL) {
2872 // check for scalar promotion
2873 Node* n = u_pk->at(0)->in(u_idx);
2874 for (uint i = 1; i < u_pk->size(); i++) {
2875 if (u_pk->at(i)->in(u_idx) != n) return false;
2876 }
2877 return true;
2878 }
2879 if (VectorNode::is_muladds2i(use)) {
2880 // MulAddS2I takes shorts and produces ints - hence the special checks
2881 // on alignment and size.
2882 if (u_pk->size() * 2 != d_pk->size()) {
2883 return false;
2884 }
2885 for (uint i = 0; i < MIN2(d_pk->size(), u_pk->size()); i++) {
2886 Node* ui = u_pk->at(i);
2887 Node* di = d_pk->at(i);
2888 if (alignment(ui) != alignment(di) * 2) {
2889 return false;
2890 }
2891 }
2892 return true;
2893 }
2894 if (u_pk->size() != d_pk->size())
2895 return false;
2896 for (uint i = 0; i < u_pk->size(); i++) {
2897 Node* ui = u_pk->at(i);
2898 Node* di = d_pk->at(i);
2899 if (ui->in(u_idx) != di || alignment(ui) != alignment(di))
2900 return false;
2901 }
2902 return true;
2903 }
2904
2905 //------------------------------construct_bb---------------------------
2906 // Construct reverse postorder list of block members
construct_bb()2907 bool SuperWord::construct_bb() {
2908 Node* entry = bb();
2909
2910 assert(_stk.length() == 0, "stk is empty");
2911 assert(_block.length() == 0, "block is empty");
2912 assert(_data_entry.length() == 0, "data_entry is empty");
2913 assert(_mem_slice_head.length() == 0, "mem_slice_head is empty");
2914 assert(_mem_slice_tail.length() == 0, "mem_slice_tail is empty");
2915
2916 // Find non-control nodes with no inputs from within block,
2917 // create a temporary map from node _idx to bb_idx for use
2918 // by the visited and post_visited sets,
2919 // and count number of nodes in block.
2920 int bb_ct = 0;
2921 for (uint i = 0; i < lpt()->_body.size(); i++) {
2922 Node *n = lpt()->_body.at(i);
2923 set_bb_idx(n, i); // Create a temporary map
2924 if (in_bb(n)) {
2925 if (n->is_LoadStore() || n->is_MergeMem() ||
2926 (n->is_Proj() && !n->as_Proj()->is_CFG())) {
2927 // Bailout if the loop has LoadStore, MergeMem or data Proj
2928 // nodes. Superword optimization does not work with them.
2929 return false;
2930 }
2931 bb_ct++;
2932 if (!n->is_CFG()) {
2933 bool found = false;
2934 for (uint j = 0; j < n->req(); j++) {
2935 Node* def = n->in(j);
2936 if (def && in_bb(def)) {
2937 found = true;
2938 break;
2939 }
2940 }
2941 if (!found) {
2942 assert(n != entry, "can't be entry");
2943 _data_entry.push(n);
2944 }
2945 }
2946 }
2947 }
2948
2949 // Find memory slices (head and tail)
2950 for (DUIterator_Fast imax, i = lp()->fast_outs(imax); i < imax; i++) {
2951 Node *n = lp()->fast_out(i);
2952 if (in_bb(n) && (n->is_Phi() && n->bottom_type() == Type::MEMORY)) {
2953 Node* n_tail = n->in(LoopNode::LoopBackControl);
2954 if (n_tail != n->in(LoopNode::EntryControl)) {
2955 if (!n_tail->is_Mem()) {
2956 assert(n_tail->is_Mem(), "unexpected node for memory slice: %s", n_tail->Name());
2957 return false; // Bailout
2958 }
2959 _mem_slice_head.push(n);
2960 _mem_slice_tail.push(n_tail);
2961 }
2962 }
2963 }
2964
2965 // Create an RPO list of nodes in block
2966
2967 visited_clear();
2968 post_visited_clear();
2969
2970 // Push all non-control nodes with no inputs from within block, then control entry
2971 for (int j = 0; j < _data_entry.length(); j++) {
2972 Node* n = _data_entry.at(j);
2973 visited_set(n);
2974 _stk.push(n);
2975 }
2976 visited_set(entry);
2977 _stk.push(entry);
2978
2979 // Do a depth first walk over out edges
2980 int rpo_idx = bb_ct - 1;
2981 int size;
2982 int reduction_uses = 0;
2983 while ((size = _stk.length()) > 0) {
2984 Node* n = _stk.top(); // Leave node on stack
2985 if (!visited_test_set(n)) {
2986 // forward arc in graph
2987 } else if (!post_visited_test(n)) {
2988 // cross or back arc
2989 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2990 Node *use = n->fast_out(i);
2991 if (in_bb(use) && !visited_test(use) &&
2992 // Don't go around backedge
2993 (!use->is_Phi() || n == entry)) {
2994 if (use->is_reduction()) {
2995 // First see if we can map the reduction on the given system we are on, then
2996 // make a data entry operation for each reduction we see.
2997 BasicType bt = use->bottom_type()->basic_type();
2998 if (ReductionNode::implemented(use->Opcode(), Matcher::min_vector_size(bt), bt)) {
2999 reduction_uses++;
3000 }
3001 }
3002 _stk.push(use);
3003 }
3004 }
3005 if (_stk.length() == size) {
3006 // There were no additional uses, post visit node now
3007 _stk.pop(); // Remove node from stack
3008 assert(rpo_idx >= 0, "");
3009 _block.at_put_grow(rpo_idx, n);
3010 rpo_idx--;
3011 post_visited_set(n);
3012 assert(rpo_idx >= 0 || _stk.is_empty(), "");
3013 }
3014 } else {
3015 _stk.pop(); // Remove post-visited node from stack
3016 }
3017 }//while
3018
3019 int ii_current = -1;
3020 unsigned int load_idx = (unsigned int)-1;
3021 _ii_order.clear();
3022 // Create real map of block indices for nodes
3023 for (int j = 0; j < _block.length(); j++) {
3024 Node* n = _block.at(j);
3025 set_bb_idx(n, j);
3026 if (_do_vector_loop && n->is_Load()) {
3027 if (ii_current == -1) {
3028 ii_current = _clone_map.gen(n->_idx);
3029 _ii_order.push(ii_current);
3030 load_idx = _clone_map.idx(n->_idx);
3031 } else if (_clone_map.idx(n->_idx) == load_idx && _clone_map.gen(n->_idx) != ii_current) {
3032 ii_current = _clone_map.gen(n->_idx);
3033 _ii_order.push(ii_current);
3034 }
3035 }
3036 }//for
3037
3038 // Ensure extra info is allocated.
3039 initialize_bb();
3040
3041 #ifndef PRODUCT
3042 if (_vector_loop_debug && _ii_order.length() > 0) {
3043 tty->print("SuperWord::construct_bb: List of generations: ");
3044 for (int jj = 0; jj < _ii_order.length(); ++jj) {
3045 tty->print(" %d:%d", jj, _ii_order.at(jj));
3046 }
3047 tty->print_cr(" ");
3048 }
3049 if (TraceSuperWord) {
3050 print_bb();
3051 tty->print_cr("\ndata entry nodes: %s", _data_entry.length() > 0 ? "" : "NONE");
3052 for (int m = 0; m < _data_entry.length(); m++) {
3053 tty->print("%3d ", m);
3054 _data_entry.at(m)->dump();
3055 }
3056 tty->print_cr("\nmemory slices: %s", _mem_slice_head.length() > 0 ? "" : "NONE");
3057 for (int m = 0; m < _mem_slice_head.length(); m++) {
3058 tty->print("%3d ", m); _mem_slice_head.at(m)->dump();
3059 tty->print(" "); _mem_slice_tail.at(m)->dump();
3060 }
3061 }
3062 #endif
3063 assert(rpo_idx == -1 && bb_ct == _block.length(), "all block members found");
3064 return (_mem_slice_head.length() > 0) || (reduction_uses > 0) || (_data_entry.length() > 0);
3065 }
3066
3067 //------------------------------initialize_bb---------------------------
3068 // Initialize per node info
initialize_bb()3069 void SuperWord::initialize_bb() {
3070 Node* last = _block.at(_block.length() - 1);
3071 grow_node_info(bb_idx(last));
3072 }
3073
3074 //------------------------------bb_insert_after---------------------------
3075 // Insert n into block after pos
bb_insert_after(Node * n,int pos)3076 void SuperWord::bb_insert_after(Node* n, int pos) {
3077 int n_pos = pos + 1;
3078 // Make room
3079 for (int i = _block.length() - 1; i >= n_pos; i--) {
3080 _block.at_put_grow(i+1, _block.at(i));
3081 }
3082 for (int j = _node_info.length() - 1; j >= n_pos; j--) {
3083 _node_info.at_put_grow(j+1, _node_info.at(j));
3084 }
3085 // Set value
3086 _block.at_put_grow(n_pos, n);
3087 _node_info.at_put_grow(n_pos, SWNodeInfo::initial);
3088 // Adjust map from node->_idx to _block index
3089 for (int i = n_pos; i < _block.length(); i++) {
3090 set_bb_idx(_block.at(i), i);
3091 }
3092 }
3093
3094 //------------------------------compute_max_depth---------------------------
3095 // Compute max depth for expressions from beginning of block
3096 // Use to prune search paths during test for independence.
compute_max_depth()3097 void SuperWord::compute_max_depth() {
3098 int ct = 0;
3099 bool again;
3100 do {
3101 again = false;
3102 for (int i = 0; i < _block.length(); i++) {
3103 Node* n = _block.at(i);
3104 if (!n->is_Phi()) {
3105 int d_orig = depth(n);
3106 int d_in = 0;
3107 for (DepPreds preds(n, _dg); !preds.done(); preds.next()) {
3108 Node* pred = preds.current();
3109 if (in_bb(pred)) {
3110 d_in = MAX2(d_in, depth(pred));
3111 }
3112 }
3113 if (d_in + 1 != d_orig) {
3114 set_depth(n, d_in + 1);
3115 again = true;
3116 }
3117 }
3118 }
3119 ct++;
3120 } while (again);
3121
3122 if (TraceSuperWord && Verbose) {
3123 tty->print_cr("compute_max_depth iterated: %d times", ct);
3124 }
3125 }
3126
3127 //-------------------------compute_vector_element_type-----------------------
3128 // Compute necessary vector element type for expressions
3129 // This propagates backwards a narrower integer type when the
3130 // upper bits of the value are not needed.
3131 // Example: char a,b,c; a = b + c;
3132 // Normally the type of the add is integer, but for packed character
3133 // operations the type of the add needs to be char.
compute_vector_element_type()3134 void SuperWord::compute_vector_element_type() {
3135 if (TraceSuperWord && Verbose) {
3136 tty->print_cr("\ncompute_velt_type:");
3137 }
3138
3139 // Initial type
3140 for (int i = 0; i < _block.length(); i++) {
3141 Node* n = _block.at(i);
3142 set_velt_type(n, container_type(n));
3143 }
3144
3145 // Propagate integer narrowed type backwards through operations
3146 // that don't depend on higher order bits
3147 for (int i = _block.length() - 1; i >= 0; i--) {
3148 Node* n = _block.at(i);
3149 // Only integer types need be examined
3150 const Type* vtn = velt_type(n);
3151 if (vtn->basic_type() == T_INT) {
3152 uint start, end;
3153 VectorNode::vector_operands(n, &start, &end);
3154
3155 for (uint j = start; j < end; j++) {
3156 Node* in = n->in(j);
3157 // Don't propagate through a memory
3158 if (!in->is_Mem() && in_bb(in) && velt_type(in)->basic_type() == T_INT &&
3159 data_size(n) < data_size(in)) {
3160 bool same_type = true;
3161 for (DUIterator_Fast kmax, k = in->fast_outs(kmax); k < kmax; k++) {
3162 Node *use = in->fast_out(k);
3163 if (!in_bb(use) || !same_velt_type(use, n)) {
3164 same_type = false;
3165 break;
3166 }
3167 }
3168 if (same_type) {
3169 // In any Java arithmetic operation, operands of small integer types
3170 // (boolean, byte, char & short) should be promoted to int first. As
3171 // vector elements of small types don't have upper bits of int, for
3172 // RShiftI or AbsI operations, the compiler has to know the precise
3173 // signedness info of the 1st operand. These operations shouldn't be
3174 // vectorized if the signedness info is imprecise.
3175 const Type* vt = vtn;
3176 int op = in->Opcode();
3177 if (VectorNode::is_shift(in) || op == Op_AbsI) {
3178 Node* load = in->in(1);
3179 if (load->is_Load() && in_bb(load) && (velt_type(load)->basic_type() == T_INT)) {
3180 // Only Load nodes distinguish signed (LoadS/LoadB) and unsigned
3181 // (LoadUS/LoadUB) values. Store nodes only have one version.
3182 vt = velt_type(load);
3183 } else if (op != Op_LShiftI) {
3184 // Widen type to int to avoid the creation of vector nodes. Note
3185 // that left shifts work regardless of the signedness.
3186 vt = TypeInt::INT;
3187 }
3188 }
3189 set_velt_type(in, vt);
3190 }
3191 }
3192 }
3193 }
3194 }
3195 #ifndef PRODUCT
3196 if (TraceSuperWord && Verbose) {
3197 for (int i = 0; i < _block.length(); i++) {
3198 Node* n = _block.at(i);
3199 velt_type(n)->dump();
3200 tty->print("\t");
3201 n->dump();
3202 }
3203 }
3204 #endif
3205 }
3206
3207 //------------------------------memory_alignment---------------------------
3208 // Alignment within a vector memory reference
memory_alignment(MemNode * s,int iv_adjust)3209 int SuperWord::memory_alignment(MemNode* s, int iv_adjust) {
3210 #ifndef PRODUCT
3211 if(TraceSuperWord && Verbose) {
3212 tty->print("SuperWord::memory_alignment within a vector memory reference for %d: ", s->_idx); s->dump();
3213 }
3214 #endif
3215 NOT_PRODUCT(SWPointer::Tracer::Depth ddd(0);)
3216 SWPointer p(s, this, NULL, false);
3217 if (!p.valid()) {
3218 NOT_PRODUCT(if(is_trace_alignment()) tty->print("SWPointer::memory_alignment: SWPointer p invalid, return bottom_align");)
3219 return bottom_align;
3220 }
3221 int vw = get_vw_bytes_special(s);
3222 if (vw < 2) {
3223 NOT_PRODUCT(if(is_trace_alignment()) tty->print_cr("SWPointer::memory_alignment: vector_width_in_bytes < 2, return bottom_align");)
3224 return bottom_align; // No vectors for this type
3225 }
3226 int offset = p.offset_in_bytes();
3227 offset += iv_adjust*p.memory_size();
3228 int off_rem = offset % vw;
3229 int off_mod = off_rem >= 0 ? off_rem : off_rem + vw;
3230 if (TraceSuperWord && Verbose) {
3231 tty->print_cr("SWPointer::memory_alignment: off_rem = %d, off_mod = %d", off_rem, off_mod);
3232 }
3233 return off_mod;
3234 }
3235
3236 //---------------------------container_type---------------------------
3237 // Smallest type containing range of values
container_type(Node * n)3238 const Type* SuperWord::container_type(Node* n) {
3239 if (n->is_Mem()) {
3240 BasicType bt = n->as_Mem()->memory_type();
3241 if (n->is_Store() && (bt == T_CHAR)) {
3242 // Use T_SHORT type instead of T_CHAR for stored values because any
3243 // preceding arithmetic operation extends values to signed Int.
3244 bt = T_SHORT;
3245 }
3246 if (n->Opcode() == Op_LoadUB) {
3247 // Adjust type for unsigned byte loads, it is important for right shifts.
3248 // T_BOOLEAN is used because there is no basic type representing type
3249 // TypeInt::UBYTE. Use of T_BOOLEAN for vectors is fine because only
3250 // size (one byte) and sign is important.
3251 bt = T_BOOLEAN;
3252 }
3253 return Type::get_const_basic_type(bt);
3254 }
3255 const Type* t = _igvn.type(n);
3256 if (t->basic_type() == T_INT) {
3257 // A narrow type of arithmetic operations will be determined by
3258 // propagating the type of memory operations.
3259 return TypeInt::INT;
3260 }
3261 return t;
3262 }
3263
same_velt_type(Node * n1,Node * n2)3264 bool SuperWord::same_velt_type(Node* n1, Node* n2) {
3265 const Type* vt1 = velt_type(n1);
3266 const Type* vt2 = velt_type(n2);
3267 if (vt1->basic_type() == T_INT && vt2->basic_type() == T_INT) {
3268 // Compare vectors element sizes for integer types.
3269 return data_size(n1) == data_size(n2);
3270 }
3271 return vt1 == vt2;
3272 }
3273
3274 //------------------------------in_packset---------------------------
3275 // Are s1 and s2 in a pack pair and ordered as s1,s2?
in_packset(Node * s1,Node * s2)3276 bool SuperWord::in_packset(Node* s1, Node* s2) {
3277 for (int i = 0; i < _packset.length(); i++) {
3278 Node_List* p = _packset.at(i);
3279 assert(p->size() == 2, "must be");
3280 if (p->at(0) == s1 && p->at(p->size()-1) == s2) {
3281 return true;
3282 }
3283 }
3284 return false;
3285 }
3286
3287 //------------------------------in_pack---------------------------
3288 // Is s in pack p?
in_pack(Node * s,Node_List * p)3289 Node_List* SuperWord::in_pack(Node* s, Node_List* p) {
3290 for (uint i = 0; i < p->size(); i++) {
3291 if (p->at(i) == s) {
3292 return p;
3293 }
3294 }
3295 return NULL;
3296 }
3297
3298 //------------------------------remove_pack_at---------------------------
3299 // Remove the pack at position pos in the packset
remove_pack_at(int pos)3300 void SuperWord::remove_pack_at(int pos) {
3301 Node_List* p = _packset.at(pos);
3302 for (uint i = 0; i < p->size(); i++) {
3303 Node* s = p->at(i);
3304 set_my_pack(s, NULL);
3305 }
3306 _packset.remove_at(pos);
3307 }
3308
packset_sort(int n)3309 void SuperWord::packset_sort(int n) {
3310 // simple bubble sort so that we capitalize with O(n) when its already sorted
3311 while (n != 0) {
3312 bool swapped = false;
3313 for (int i = 1; i < n; i++) {
3314 Node_List* q_low = _packset.at(i-1);
3315 Node_List* q_i = _packset.at(i);
3316
3317 // only swap when we find something to swap
3318 if (alignment(q_low->at(0)) > alignment(q_i->at(0))) {
3319 Node_List* t = q_i;
3320 *(_packset.adr_at(i)) = q_low;
3321 *(_packset.adr_at(i-1)) = q_i;
3322 swapped = true;
3323 }
3324 }
3325 if (swapped == false) break;
3326 n--;
3327 }
3328 }
3329
3330 //------------------------------executed_first---------------------------
3331 // Return the node executed first in pack p. Uses the RPO block list
3332 // to determine order.
executed_first(Node_List * p)3333 Node* SuperWord::executed_first(Node_List* p) {
3334 Node* n = p->at(0);
3335 int n_rpo = bb_idx(n);
3336 for (uint i = 1; i < p->size(); i++) {
3337 Node* s = p->at(i);
3338 int s_rpo = bb_idx(s);
3339 if (s_rpo < n_rpo) {
3340 n = s;
3341 n_rpo = s_rpo;
3342 }
3343 }
3344 return n;
3345 }
3346
3347 //------------------------------executed_last---------------------------
3348 // Return the node executed last in pack p.
executed_last(Node_List * p)3349 Node* SuperWord::executed_last(Node_List* p) {
3350 Node* n = p->at(0);
3351 int n_rpo = bb_idx(n);
3352 for (uint i = 1; i < p->size(); i++) {
3353 Node* s = p->at(i);
3354 int s_rpo = bb_idx(s);
3355 if (s_rpo > n_rpo) {
3356 n = s;
3357 n_rpo = s_rpo;
3358 }
3359 }
3360 return n;
3361 }
3362
control_dependency(Node_List * p)3363 LoadNode::ControlDependency SuperWord::control_dependency(Node_List* p) {
3364 LoadNode::ControlDependency dep = LoadNode::DependsOnlyOnTest;
3365 for (uint i = 0; i < p->size(); i++) {
3366 Node* n = p->at(i);
3367 assert(n->is_Load(), "only meaningful for loads");
3368 if (!n->depends_only_on_test()) {
3369 if (n->as_Load()->has_unknown_control_dependency() &&
3370 dep != LoadNode::Pinned) {
3371 // Upgrade to unknown control...
3372 dep = LoadNode::UnknownControl;
3373 } else {
3374 // Otherwise, we must pin it.
3375 dep = LoadNode::Pinned;
3376 }
3377 }
3378 }
3379 return dep;
3380 }
3381
3382
3383 //----------------------------align_initial_loop_index---------------------------
3384 // Adjust pre-loop limit so that in main loop, a load/store reference
3385 // to align_to_ref will be a position zero in the vector.
3386 // (iv + k) mod vector_align == 0
align_initial_loop_index(MemNode * align_to_ref)3387 void SuperWord::align_initial_loop_index(MemNode* align_to_ref) {
3388 CountedLoopNode *main_head = lp()->as_CountedLoop();
3389 assert(main_head->is_main_loop(), "");
3390 CountedLoopEndNode* pre_end = get_pre_loop_end(main_head);
3391 assert(pre_end != NULL, "we must have a correct pre-loop");
3392 Node *pre_opaq1 = pre_end->limit();
3393 assert(pre_opaq1->Opcode() == Op_Opaque1, "");
3394 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
3395 Node *lim0 = pre_opaq->in(1);
3396
3397 // Where we put new limit calculations
3398 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
3399
3400 // Ensure the original loop limit is available from the
3401 // pre-loop Opaque1 node.
3402 Node *orig_limit = pre_opaq->original_loop_limit();
3403 assert(orig_limit != NULL && _igvn.type(orig_limit) != Type::TOP, "");
3404
3405 SWPointer align_to_ref_p(align_to_ref, this, NULL, false);
3406 assert(align_to_ref_p.valid(), "sanity");
3407
3408 // Given:
3409 // lim0 == original pre loop limit
3410 // V == v_align (power of 2)
3411 // invar == extra invariant piece of the address expression
3412 // e == offset [ +/- invar ]
3413 //
3414 // When reassociating expressions involving '%' the basic rules are:
3415 // (a - b) % k == 0 => a % k == b % k
3416 // and:
3417 // (a + b) % k == 0 => a % k == (k - b) % k
3418 //
3419 // For stride > 0 && scale > 0,
3420 // Derive the new pre-loop limit "lim" such that the two constraints:
3421 // (1) lim = lim0 + N (where N is some positive integer < V)
3422 // (2) (e + lim) % V == 0
3423 // are true.
3424 //
3425 // Substituting (1) into (2),
3426 // (e + lim0 + N) % V == 0
3427 // solve for N:
3428 // N = (V - (e + lim0)) % V
3429 // substitute back into (1), so that new limit
3430 // lim = lim0 + (V - (e + lim0)) % V
3431 //
3432 // For stride > 0 && scale < 0
3433 // Constraints:
3434 // lim = lim0 + N
3435 // (e - lim) % V == 0
3436 // Solving for lim:
3437 // (e - lim0 - N) % V == 0
3438 // N = (e - lim0) % V
3439 // lim = lim0 + (e - lim0) % V
3440 //
3441 // For stride < 0 && scale > 0
3442 // Constraints:
3443 // lim = lim0 - N
3444 // (e + lim) % V == 0
3445 // Solving for lim:
3446 // (e + lim0 - N) % V == 0
3447 // N = (e + lim0) % V
3448 // lim = lim0 - (e + lim0) % V
3449 //
3450 // For stride < 0 && scale < 0
3451 // Constraints:
3452 // lim = lim0 - N
3453 // (e - lim) % V == 0
3454 // Solving for lim:
3455 // (e - lim0 + N) % V == 0
3456 // N = (V - (e - lim0)) % V
3457 // lim = lim0 - (V - (e - lim0)) % V
3458
3459 int vw = vector_width_in_bytes(align_to_ref);
3460 int stride = iv_stride();
3461 int scale = align_to_ref_p.scale_in_bytes();
3462 int elt_size = align_to_ref_p.memory_size();
3463 int v_align = vw / elt_size;
3464 assert(v_align > 1, "sanity");
3465 int offset = align_to_ref_p.offset_in_bytes() / elt_size;
3466 Node *offsn = _igvn.intcon(offset);
3467
3468 Node *e = offsn;
3469 if (align_to_ref_p.invar() != NULL) {
3470 // incorporate any extra invariant piece producing (offset +/- invar) >>> log2(elt)
3471 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3472 Node* invar = align_to_ref_p.invar();
3473 if (_igvn.type(invar)->isa_long()) {
3474 // Computations are done % (vector width/element size) so it's
3475 // safe to simply convert invar to an int and loose the upper 32
3476 // bit half.
3477 invar = new ConvL2INode(invar);
3478 _igvn.register_new_node_with_optimizer(invar);
3479 }
3480 Node* aref = new URShiftINode(invar, log2_elt);
3481 _igvn.register_new_node_with_optimizer(aref);
3482 _phase->set_ctrl(aref, pre_ctrl);
3483 if (align_to_ref_p.negate_invar()) {
3484 e = new SubINode(e, aref);
3485 } else {
3486 e = new AddINode(e, aref);
3487 }
3488 _igvn.register_new_node_with_optimizer(e);
3489 _phase->set_ctrl(e, pre_ctrl);
3490 }
3491 if (vw > ObjectAlignmentInBytes || align_to_ref_p.base()->is_top()) {
3492 // incorporate base e +/- base && Mask >>> log2(elt)
3493 Node* xbase = new CastP2XNode(NULL, align_to_ref_p.adr());
3494 _igvn.register_new_node_with_optimizer(xbase);
3495 #ifdef _LP64
3496 xbase = new ConvL2INode(xbase);
3497 _igvn.register_new_node_with_optimizer(xbase);
3498 #endif
3499 Node* mask = _igvn.intcon(vw-1);
3500 Node* masked_xbase = new AndINode(xbase, mask);
3501 _igvn.register_new_node_with_optimizer(masked_xbase);
3502 Node* log2_elt = _igvn.intcon(exact_log2(elt_size));
3503 Node* bref = new URShiftINode(masked_xbase, log2_elt);
3504 _igvn.register_new_node_with_optimizer(bref);
3505 _phase->set_ctrl(bref, pre_ctrl);
3506 e = new AddINode(e, bref);
3507 _igvn.register_new_node_with_optimizer(e);
3508 _phase->set_ctrl(e, pre_ctrl);
3509 }
3510
3511 // compute e +/- lim0
3512 if (scale < 0) {
3513 e = new SubINode(e, lim0);
3514 } else {
3515 e = new AddINode(e, lim0);
3516 }
3517 _igvn.register_new_node_with_optimizer(e);
3518 _phase->set_ctrl(e, pre_ctrl);
3519
3520 if (stride * scale > 0) {
3521 // compute V - (e +/- lim0)
3522 Node* va = _igvn.intcon(v_align);
3523 e = new SubINode(va, e);
3524 _igvn.register_new_node_with_optimizer(e);
3525 _phase->set_ctrl(e, pre_ctrl);
3526 }
3527 // compute N = (exp) % V
3528 Node* va_msk = _igvn.intcon(v_align - 1);
3529 Node* N = new AndINode(e, va_msk);
3530 _igvn.register_new_node_with_optimizer(N);
3531 _phase->set_ctrl(N, pre_ctrl);
3532
3533 // substitute back into (1), so that new limit
3534 // lim = lim0 + N
3535 Node* lim;
3536 if (stride < 0) {
3537 lim = new SubINode(lim0, N);
3538 } else {
3539 lim = new AddINode(lim0, N);
3540 }
3541 _igvn.register_new_node_with_optimizer(lim);
3542 _phase->set_ctrl(lim, pre_ctrl);
3543 Node* constrained =
3544 (stride > 0) ? (Node*) new MinINode(lim, orig_limit)
3545 : (Node*) new MaxINode(lim, orig_limit);
3546 _igvn.register_new_node_with_optimizer(constrained);
3547 _phase->set_ctrl(constrained, pre_ctrl);
3548 _igvn.replace_input_of(pre_opaq, 1, constrained);
3549 }
3550
3551 //----------------------------get_pre_loop_end---------------------------
3552 // Find pre loop end from main loop. Returns null if none.
get_pre_loop_end(CountedLoopNode * cl)3553 CountedLoopEndNode* SuperWord::get_pre_loop_end(CountedLoopNode* cl) {
3554 // The loop cannot be optimized if the graph shape at
3555 // the loop entry is inappropriate.
3556 if (!PhaseIdealLoop::is_canonical_loop_entry(cl)) {
3557 return NULL;
3558 }
3559
3560 Node* p_f = cl->skip_predicates()->in(0)->in(0);
3561 if (!p_f->is_IfFalse()) return NULL;
3562 if (!p_f->in(0)->is_CountedLoopEnd()) return NULL;
3563 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
3564 CountedLoopNode* loop_node = pre_end->loopnode();
3565 if (loop_node == NULL || !loop_node->is_pre_loop()) return NULL;
3566 return pre_end;
3567 }
3568
3569 //------------------------------init---------------------------
init()3570 void SuperWord::init() {
3571 _dg.init();
3572 _packset.clear();
3573 _disjoint_ptrs.clear();
3574 _block.clear();
3575 _post_block.clear();
3576 _data_entry.clear();
3577 _mem_slice_head.clear();
3578 _mem_slice_tail.clear();
3579 _iteration_first.clear();
3580 _iteration_last.clear();
3581 _node_info.clear();
3582 _align_to_ref = NULL;
3583 _lpt = NULL;
3584 _lp = NULL;
3585 _bb = NULL;
3586 _iv = NULL;
3587 _race_possible = 0;
3588 _early_return = false;
3589 _num_work_vecs = 0;
3590 _num_reductions = 0;
3591 }
3592
3593 //------------------------------restart---------------------------
restart()3594 void SuperWord::restart() {
3595 _dg.init();
3596 _packset.clear();
3597 _disjoint_ptrs.clear();
3598 _block.clear();
3599 _post_block.clear();
3600 _data_entry.clear();
3601 _mem_slice_head.clear();
3602 _mem_slice_tail.clear();
3603 _node_info.clear();
3604 }
3605
3606 //------------------------------print_packset---------------------------
print_packset()3607 void SuperWord::print_packset() {
3608 #ifndef PRODUCT
3609 tty->print_cr("packset");
3610 for (int i = 0; i < _packset.length(); i++) {
3611 tty->print_cr("Pack: %d", i);
3612 Node_List* p = _packset.at(i);
3613 print_pack(p);
3614 }
3615 #endif
3616 }
3617
3618 //------------------------------print_pack---------------------------
print_pack(Node_List * p)3619 void SuperWord::print_pack(Node_List* p) {
3620 for (uint i = 0; i < p->size(); i++) {
3621 print_stmt(p->at(i));
3622 }
3623 }
3624
3625 //------------------------------print_bb---------------------------
print_bb()3626 void SuperWord::print_bb() {
3627 #ifndef PRODUCT
3628 tty->print_cr("\nBlock");
3629 for (int i = 0; i < _block.length(); i++) {
3630 Node* n = _block.at(i);
3631 tty->print("%d ", i);
3632 if (n) {
3633 n->dump();
3634 }
3635 }
3636 #endif
3637 }
3638
3639 //------------------------------print_stmt---------------------------
print_stmt(Node * s)3640 void SuperWord::print_stmt(Node* s) {
3641 #ifndef PRODUCT
3642 tty->print(" align: %d \t", alignment(s));
3643 s->dump();
3644 #endif
3645 }
3646
3647 //------------------------------blank---------------------------
blank(uint depth)3648 char* SuperWord::blank(uint depth) {
3649 static char blanks[101];
3650 assert(depth < 101, "too deep");
3651 for (uint i = 0; i < depth; i++) blanks[i] = ' ';
3652 blanks[depth] = '\0';
3653 return blanks;
3654 }
3655
3656
3657 //==============================SWPointer===========================
3658 #ifndef PRODUCT
3659 int SWPointer::Tracer::_depth = 0;
3660 #endif
3661 //----------------------------SWPointer------------------------
SWPointer(MemNode * mem,SuperWord * slp,Node_Stack * nstack,bool analyze_only)3662 SWPointer::SWPointer(MemNode* mem, SuperWord* slp, Node_Stack *nstack, bool analyze_only) :
3663 _mem(mem), _slp(slp), _base(NULL), _adr(NULL),
3664 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3665 _nstack(nstack), _analyze_only(analyze_only),
3666 _stack_idx(0)
3667 #ifndef PRODUCT
3668 , _tracer(slp)
3669 #endif
3670 {
3671 NOT_PRODUCT(_tracer.ctor_1(mem);)
3672
3673 Node* adr = mem->in(MemNode::Address);
3674 if (!adr->is_AddP()) {
3675 assert(!valid(), "too complex");
3676 return;
3677 }
3678 // Match AddP(base, AddP(ptr, k*iv [+ invariant]), constant)
3679 Node* base = adr->in(AddPNode::Base);
3680 // The base address should be loop invariant
3681 if (!invariant(base)) {
3682 assert(!valid(), "base address is loop variant");
3683 return;
3684 }
3685 // unsafe references require misaligned vector access support
3686 if (base->is_top() && !Matcher::misaligned_vectors_ok()) {
3687 assert(!valid(), "unsafe access");
3688 return;
3689 }
3690
3691 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.store_depth();)
3692 NOT_PRODUCT(_tracer.ctor_2(adr);)
3693
3694 int i;
3695 for (i = 0; i < 3; i++) {
3696 NOT_PRODUCT(_tracer.ctor_3(adr, i);)
3697
3698 if (!scaled_iv_plus_offset(adr->in(AddPNode::Offset))) {
3699 assert(!valid(), "too complex");
3700 return;
3701 }
3702 adr = adr->in(AddPNode::Address);
3703 NOT_PRODUCT(_tracer.ctor_4(adr, i);)
3704
3705 if (base == adr || !adr->is_AddP()) {
3706 NOT_PRODUCT(_tracer.ctor_5(adr, base, i);)
3707 break; // stop looking at addp's
3708 }
3709 }
3710 if (!invariant(adr)) {
3711 assert(!valid(), "adr is loop variant");
3712 return;
3713 }
3714
3715 if (!base->is_top() && adr != base) {
3716 assert(!valid(), "adr and base differ");
3717 return;
3718 }
3719
3720 NOT_PRODUCT(if(_slp->is_trace_alignment()) _tracer.restore_depth();)
3721 NOT_PRODUCT(_tracer.ctor_6(mem);)
3722
3723 _base = base;
3724 _adr = adr;
3725 assert(valid(), "Usable");
3726 }
3727
3728 // Following is used to create a temporary object during
3729 // the pattern match of an address expression.
SWPointer(SWPointer * p)3730 SWPointer::SWPointer(SWPointer* p) :
3731 _mem(p->_mem), _slp(p->_slp), _base(NULL), _adr(NULL),
3732 _scale(0), _offset(0), _invar(NULL), _negate_invar(false),
3733 _nstack(p->_nstack), _analyze_only(p->_analyze_only),
3734 _stack_idx(p->_stack_idx)
3735 #ifndef PRODUCT
3736 , _tracer(p->_slp)
3737 #endif
3738 {}
3739
3740
invariant(Node * n)3741 bool SWPointer::invariant(Node* n) {
3742 NOT_PRODUCT(Tracer::Depth dd;)
3743 Node *n_c = phase()->get_ctrl(n);
3744 NOT_PRODUCT(_tracer.invariant_1(n, n_c);)
3745 return !lpt()->is_member(phase()->get_loop(n_c));
3746 }
3747 //------------------------scaled_iv_plus_offset--------------------
3748 // Match: k*iv + offset
3749 // where: k is a constant that maybe zero, and
3750 // offset is (k2 [+/- invariant]) where k2 maybe zero and invariant is optional
scaled_iv_plus_offset(Node * n)3751 bool SWPointer::scaled_iv_plus_offset(Node* n) {
3752 NOT_PRODUCT(Tracer::Depth ddd;)
3753 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_1(n);)
3754
3755 if (scaled_iv(n)) {
3756 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_2(n);)
3757 return true;
3758 }
3759
3760 if (offset_plus_k(n)) {
3761 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_3(n);)
3762 return true;
3763 }
3764
3765 int opc = n->Opcode();
3766 if (opc == Op_AddI) {
3767 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2))) {
3768 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_4(n);)
3769 return true;
3770 }
3771 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3772 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_5(n);)
3773 return true;
3774 }
3775 } else if (opc == Op_SubI) {
3776 if (scaled_iv(n->in(1)) && offset_plus_k(n->in(2), true)) {
3777 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_6(n);)
3778 return true;
3779 }
3780 if (scaled_iv(n->in(2)) && offset_plus_k(n->in(1))) {
3781 _scale *= -1;
3782 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_7(n);)
3783 return true;
3784 }
3785 }
3786
3787 NOT_PRODUCT(_tracer.scaled_iv_plus_offset_8(n);)
3788 return false;
3789 }
3790
3791 //----------------------------scaled_iv------------------------
3792 // Match: k*iv where k is a constant that's not zero
scaled_iv(Node * n)3793 bool SWPointer::scaled_iv(Node* n) {
3794 NOT_PRODUCT(Tracer::Depth ddd;)
3795 NOT_PRODUCT(_tracer.scaled_iv_1(n);)
3796
3797 if (_scale != 0) { // already found a scale
3798 NOT_PRODUCT(_tracer.scaled_iv_2(n, _scale);)
3799 return false;
3800 }
3801
3802 if (n == iv()) {
3803 _scale = 1;
3804 NOT_PRODUCT(_tracer.scaled_iv_3(n, _scale);)
3805 return true;
3806 }
3807 if (_analyze_only && (invariant(n) == false)) {
3808 _nstack->push(n, _stack_idx++);
3809 }
3810
3811 int opc = n->Opcode();
3812 if (opc == Op_MulI) {
3813 if (n->in(1) == iv() && n->in(2)->is_Con()) {
3814 _scale = n->in(2)->get_int();
3815 NOT_PRODUCT(_tracer.scaled_iv_4(n, _scale);)
3816 return true;
3817 } else if (n->in(2) == iv() && n->in(1)->is_Con()) {
3818 _scale = n->in(1)->get_int();
3819 NOT_PRODUCT(_tracer.scaled_iv_5(n, _scale);)
3820 return true;
3821 }
3822 } else if (opc == Op_LShiftI) {
3823 if (n->in(1) == iv() && n->in(2)->is_Con()) {
3824 _scale = 1 << n->in(2)->get_int();
3825 NOT_PRODUCT(_tracer.scaled_iv_6(n, _scale);)
3826 return true;
3827 }
3828 } else if (opc == Op_ConvI2L) {
3829 if (n->in(1)->Opcode() == Op_CastII &&
3830 n->in(1)->as_CastII()->has_range_check()) {
3831 // Skip range check dependent CastII nodes
3832 n = n->in(1);
3833 }
3834 if (scaled_iv_plus_offset(n->in(1))) {
3835 NOT_PRODUCT(_tracer.scaled_iv_7(n);)
3836 return true;
3837 }
3838 } else if (opc == Op_LShiftL) {
3839 if (!has_iv() && _invar == NULL) {
3840 // Need to preserve the current _offset value, so
3841 // create a temporary object for this expression subtree.
3842 // Hacky, so should re-engineer the address pattern match.
3843 NOT_PRODUCT(Tracer::Depth dddd;)
3844 SWPointer tmp(this);
3845 NOT_PRODUCT(_tracer.scaled_iv_8(n, &tmp);)
3846
3847 if (tmp.scaled_iv_plus_offset(n->in(1))) {
3848 if (tmp._invar == NULL || _slp->do_vector_loop()) {
3849 int mult = 1 << n->in(2)->get_int();
3850 _scale = tmp._scale * mult;
3851 _offset += tmp._offset * mult;
3852 NOT_PRODUCT(_tracer.scaled_iv_9(n, _scale, _offset, mult);)
3853 return true;
3854 }
3855 }
3856 }
3857 }
3858 NOT_PRODUCT(_tracer.scaled_iv_10(n);)
3859 return false;
3860 }
3861
3862 //----------------------------offset_plus_k------------------------
3863 // Match: offset is (k [+/- invariant])
3864 // where k maybe zero and invariant is optional, but not both.
offset_plus_k(Node * n,bool negate)3865 bool SWPointer::offset_plus_k(Node* n, bool negate) {
3866 NOT_PRODUCT(Tracer::Depth ddd;)
3867 NOT_PRODUCT(_tracer.offset_plus_k_1(n);)
3868
3869 int opc = n->Opcode();
3870 if (opc == Op_ConI) {
3871 _offset += negate ? -(n->get_int()) : n->get_int();
3872 NOT_PRODUCT(_tracer.offset_plus_k_2(n, _offset);)
3873 return true;
3874 } else if (opc == Op_ConL) {
3875 // Okay if value fits into an int
3876 const TypeLong* t = n->find_long_type();
3877 if (t->higher_equal(TypeLong::INT)) {
3878 jlong loff = n->get_long();
3879 jint off = (jint)loff;
3880 _offset += negate ? -off : loff;
3881 NOT_PRODUCT(_tracer.offset_plus_k_3(n, _offset);)
3882 return true;
3883 }
3884 NOT_PRODUCT(_tracer.offset_plus_k_4(n);)
3885 return false;
3886 }
3887 if (_invar != NULL) { // already has an invariant
3888 NOT_PRODUCT(_tracer.offset_plus_k_5(n, _invar);)
3889 return false;
3890 }
3891
3892 if (_analyze_only && (invariant(n) == false)) {
3893 _nstack->push(n, _stack_idx++);
3894 }
3895 if (opc == Op_AddI) {
3896 if (n->in(2)->is_Con() && invariant(n->in(1))) {
3897 _negate_invar = negate;
3898 _invar = n->in(1);
3899 _offset += negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3900 NOT_PRODUCT(_tracer.offset_plus_k_6(n, _invar, _negate_invar, _offset);)
3901 return true;
3902 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3903 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3904 _negate_invar = negate;
3905 _invar = n->in(2);
3906 NOT_PRODUCT(_tracer.offset_plus_k_7(n, _invar, _negate_invar, _offset);)
3907 return true;
3908 }
3909 }
3910 if (opc == Op_SubI) {
3911 if (n->in(2)->is_Con() && invariant(n->in(1))) {
3912 _negate_invar = negate;
3913 _invar = n->in(1);
3914 _offset += !negate ? -(n->in(2)->get_int()) : n->in(2)->get_int();
3915 NOT_PRODUCT(_tracer.offset_plus_k_8(n, _invar, _negate_invar, _offset);)
3916 return true;
3917 } else if (n->in(1)->is_Con() && invariant(n->in(2))) {
3918 _offset += negate ? -(n->in(1)->get_int()) : n->in(1)->get_int();
3919 _negate_invar = !negate;
3920 _invar = n->in(2);
3921 NOT_PRODUCT(_tracer.offset_plus_k_9(n, _invar, _negate_invar, _offset);)
3922 return true;
3923 }
3924 }
3925 if (invariant(n)) {
3926 if (opc == Op_ConvI2L) {
3927 n = n->in(1);
3928 if (n->Opcode() == Op_CastII &&
3929 n->as_CastII()->has_range_check()) {
3930 // Skip range check dependent CastII nodes
3931 assert(invariant(n), "sanity");
3932 n = n->in(1);
3933 }
3934 }
3935 _negate_invar = negate;
3936 _invar = n;
3937 NOT_PRODUCT(_tracer.offset_plus_k_10(n, _invar, _negate_invar, _offset);)
3938 return true;
3939 }
3940
3941 NOT_PRODUCT(_tracer.offset_plus_k_11(n);)
3942 return false;
3943 }
3944
3945 //----------------------------print------------------------
print()3946 void SWPointer::print() {
3947 #ifndef PRODUCT
3948 tty->print("base: %d adr: %d scale: %d offset: %d invar: %c%d\n",
3949 _base != NULL ? _base->_idx : 0,
3950 _adr != NULL ? _adr->_idx : 0,
3951 _scale, _offset,
3952 _negate_invar?'-':'+',
3953 _invar != NULL ? _invar->_idx : 0);
3954 #endif
3955 }
3956
3957 //----------------------------tracing------------------------
3958 #ifndef PRODUCT
print_depth()3959 void SWPointer::Tracer::print_depth() {
3960 for (int ii = 0; ii<_depth; ++ii) tty->print(" ");
3961 }
3962
ctor_1(Node * mem)3963 void SWPointer::Tracer::ctor_1 (Node* mem) {
3964 if(_slp->is_trace_alignment()) {
3965 print_depth(); tty->print(" %d SWPointer::SWPointer: start alignment analysis", mem->_idx); mem->dump();
3966 }
3967 }
3968
ctor_2(Node * adr)3969 void SWPointer::Tracer::ctor_2(Node* adr) {
3970 if(_slp->is_trace_alignment()) {
3971 //store_depth();
3972 inc_depth();
3973 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: ", adr->_idx); adr->dump();
3974 inc_depth();
3975 print_depth(); tty->print(" %d (base) SWPointer::SWPointer: ", adr->in(AddPNode::Base)->_idx); adr->in(AddPNode::Base)->dump();
3976 }
3977 }
3978
ctor_3(Node * adr,int i)3979 void SWPointer::Tracer::ctor_3(Node* adr, int i) {
3980 if(_slp->is_trace_alignment()) {
3981 inc_depth();
3982 Node* offset = adr->in(AddPNode::Offset);
3983 print_depth(); tty->print(" %d (offset) SWPointer::SWPointer: i = %d: ", offset->_idx, i); offset->dump();
3984 }
3985 }
3986
ctor_4(Node * adr,int i)3987 void SWPointer::Tracer::ctor_4(Node* adr, int i) {
3988 if(_slp->is_trace_alignment()) {
3989 inc_depth();
3990 print_depth(); tty->print(" %d (adr) SWPointer::SWPointer: i = %d: ", adr->_idx, i); adr->dump();
3991 }
3992 }
3993
ctor_5(Node * adr,Node * base,int i)3994 void SWPointer::Tracer::ctor_5(Node* adr, Node* base, int i) {
3995 if(_slp->is_trace_alignment()) {
3996 inc_depth();
3997 if (base == adr) {
3998 print_depth(); tty->print_cr(" \\ %d (adr) == %d (base) SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, base->_idx, i);
3999 } else if (!adr->is_AddP()) {
4000 print_depth(); tty->print_cr(" \\ %d (adr) is NOT Addp SWPointer::SWPointer: breaking analysis at i = %d", adr->_idx, i);
4001 }
4002 }
4003 }
4004
ctor_6(Node * mem)4005 void SWPointer::Tracer::ctor_6(Node* mem) {
4006 if(_slp->is_trace_alignment()) {
4007 //restore_depth();
4008 print_depth(); tty->print_cr(" %d (adr) SWPointer::SWPointer: stop analysis", mem->_idx);
4009 }
4010 }
4011
invariant_1(Node * n,Node * n_c)4012 void SWPointer::Tracer::invariant_1(Node *n, Node *n_c) {
4013 if (_slp->do_vector_loop() && _slp->is_debug() && _slp->_lpt->is_member(_slp->_phase->get_loop(n_c)) != (int)_slp->in_bb(n)) {
4014 int is_member = _slp->_lpt->is_member(_slp->_phase->get_loop(n_c));
4015 int in_bb = _slp->in_bb(n);
4016 print_depth(); tty->print(" \\ "); tty->print_cr(" %d SWPointer::invariant conditions differ: n_c %d", n->_idx, n_c->_idx);
4017 print_depth(); tty->print(" \\ "); tty->print_cr("is_member %d, in_bb %d", is_member, in_bb);
4018 print_depth(); tty->print(" \\ "); n->dump();
4019 print_depth(); tty->print(" \\ "); n_c->dump();
4020 }
4021 }
4022
scaled_iv_plus_offset_1(Node * n)4023 void SWPointer::Tracer::scaled_iv_plus_offset_1(Node* n) {
4024 if(_slp->is_trace_alignment()) {
4025 print_depth(); tty->print(" %d SWPointer::scaled_iv_plus_offset testing node: ", n->_idx);
4026 n->dump();
4027 }
4028 }
4029
scaled_iv_plus_offset_2(Node * n)4030 void SWPointer::Tracer::scaled_iv_plus_offset_2(Node* n) {
4031 if(_slp->is_trace_alignment()) {
4032 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
4033 }
4034 }
4035
scaled_iv_plus_offset_3(Node * n)4036 void SWPointer::Tracer::scaled_iv_plus_offset_3(Node* n) {
4037 if(_slp->is_trace_alignment()) {
4038 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: PASSED", n->_idx);
4039 }
4040 }
4041
scaled_iv_plus_offset_4(Node * n)4042 void SWPointer::Tracer::scaled_iv_plus_offset_4(Node* n) {
4043 if(_slp->is_trace_alignment()) {
4044 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
4045 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
4046 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
4047 }
4048 }
4049
scaled_iv_plus_offset_5(Node * n)4050 void SWPointer::Tracer::scaled_iv_plus_offset_5(Node* n) {
4051 if(_slp->is_trace_alignment()) {
4052 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_AddI PASSED", n->_idx);
4053 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
4054 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
4055 }
4056 }
4057
scaled_iv_plus_offset_6(Node * n)4058 void SWPointer::Tracer::scaled_iv_plus_offset_6(Node* n) {
4059 if(_slp->is_trace_alignment()) {
4060 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
4061 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is scaled_iv: ", n->in(1)->_idx); n->in(1)->dump();
4062 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is offset_plus_k: ", n->in(2)->_idx); n->in(2)->dump();
4063 }
4064 }
4065
scaled_iv_plus_offset_7(Node * n)4066 void SWPointer::Tracer::scaled_iv_plus_offset_7(Node* n) {
4067 if(_slp->is_trace_alignment()) {
4068 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: Op_SubI PASSED", n->_idx);
4069 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(2) is scaled_iv: ", n->in(2)->_idx); n->in(2)->dump();
4070 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv_plus_offset: in(1) is offset_plus_k: ", n->in(1)->_idx); n->in(1)->dump();
4071 }
4072 }
4073
scaled_iv_plus_offset_8(Node * n)4074 void SWPointer::Tracer::scaled_iv_plus_offset_8(Node* n) {
4075 if(_slp->is_trace_alignment()) {
4076 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv_plus_offset: FAILED", n->_idx);
4077 }
4078 }
4079
scaled_iv_1(Node * n)4080 void SWPointer::Tracer::scaled_iv_1(Node* n) {
4081 if(_slp->is_trace_alignment()) {
4082 print_depth(); tty->print(" %d SWPointer::scaled_iv: testing node: ", n->_idx); n->dump();
4083 }
4084 }
4085
scaled_iv_2(Node * n,int scale)4086 void SWPointer::Tracer::scaled_iv_2(Node* n, int scale) {
4087 if(_slp->is_trace_alignment()) {
4088 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED since another _scale has been detected before", n->_idx);
4089 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: _scale (%d) != 0", scale);
4090 }
4091 }
4092
scaled_iv_3(Node * n,int scale)4093 void SWPointer::Tracer::scaled_iv_3(Node* n, int scale) {
4094 if(_slp->is_trace_alignment()) {
4095 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: is iv, setting _scale = %d", n->_idx, scale);
4096 }
4097 }
4098
scaled_iv_4(Node * n,int scale)4099 void SWPointer::Tracer::scaled_iv_4(Node* n, int scale) {
4100 if(_slp->is_trace_alignment()) {
4101 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
4102 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
4103 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4104 }
4105 }
4106
scaled_iv_5(Node * n,int scale)4107 void SWPointer::Tracer::scaled_iv_5(Node* n, int scale) {
4108 if(_slp->is_trace_alignment()) {
4109 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_MulI PASSED, setting _scale = %d", n->_idx, scale);
4110 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is iv: ", n->in(2)->_idx); n->in(2)->dump();
4111 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4112 }
4113 }
4114
scaled_iv_6(Node * n,int scale)4115 void SWPointer::Tracer::scaled_iv_6(Node* n, int scale) {
4116 if(_slp->is_trace_alignment()) {
4117 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftI PASSED, setting _scale = %d", n->_idx, scale);
4118 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(1) is iv: ", n->in(1)->_idx); n->in(1)->dump();
4119 print_depth(); tty->print(" \\ %d SWPointer::scaled_iv: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4120 }
4121 }
4122
scaled_iv_7(Node * n)4123 void SWPointer::Tracer::scaled_iv_7(Node* n) {
4124 if(_slp->is_trace_alignment()) {
4125 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_ConvI2L PASSED", n->_idx);
4126 print_depth(); tty->print_cr(" \\ SWPointer::scaled_iv: in(1) %d is scaled_iv_plus_offset: ", n->in(1)->_idx);
4127 inc_depth(); inc_depth();
4128 print_depth(); n->in(1)->dump();
4129 dec_depth(); dec_depth();
4130 }
4131 }
4132
scaled_iv_8(Node * n,SWPointer * tmp)4133 void SWPointer::Tracer::scaled_iv_8(Node* n, SWPointer* tmp) {
4134 if(_slp->is_trace_alignment()) {
4135 print_depth(); tty->print(" %d SWPointer::scaled_iv: Op_LShiftL, creating tmp SWPointer: ", n->_idx); tmp->print();
4136 }
4137 }
4138
scaled_iv_9(Node * n,int scale,int _offset,int mult)4139 void SWPointer::Tracer::scaled_iv_9(Node* n, int scale, int _offset, int mult) {
4140 if(_slp->is_trace_alignment()) {
4141 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: Op_LShiftL PASSED, setting _scale = %d, _offset = %d", n->_idx, scale, _offset);
4142 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",
4143 n->in(1)->_idx, n->in(2)->_idx, mult, scale, _offset);
4144 inc_depth(); inc_depth();
4145 print_depth(); n->in(1)->dump();
4146 print_depth(); n->in(2)->dump();
4147 dec_depth(); dec_depth();
4148 }
4149 }
4150
scaled_iv_10(Node * n)4151 void SWPointer::Tracer::scaled_iv_10(Node* n) {
4152 if(_slp->is_trace_alignment()) {
4153 print_depth(); tty->print_cr(" %d SWPointer::scaled_iv: FAILED", n->_idx);
4154 }
4155 }
4156
offset_plus_k_1(Node * n)4157 void SWPointer::Tracer::offset_plus_k_1(Node* n) {
4158 if(_slp->is_trace_alignment()) {
4159 print_depth(); tty->print(" %d SWPointer::offset_plus_k: testing node: ", n->_idx); n->dump();
4160 }
4161 }
4162
offset_plus_k_2(Node * n,int _offset)4163 void SWPointer::Tracer::offset_plus_k_2(Node* n, int _offset) {
4164 if(_slp->is_trace_alignment()) {
4165 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConI PASSED, setting _offset = %d", n->_idx, _offset);
4166 }
4167 }
4168
offset_plus_k_3(Node * n,int _offset)4169 void SWPointer::Tracer::offset_plus_k_3(Node* n, int _offset) {
4170 if(_slp->is_trace_alignment()) {
4171 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_ConL PASSED, setting _offset = %d", n->_idx, _offset);
4172 }
4173 }
4174
offset_plus_k_4(Node * n)4175 void SWPointer::Tracer::offset_plus_k_4(Node* n) {
4176 if(_slp->is_trace_alignment()) {
4177 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4178 print_depth(); tty->print_cr(" \\ " JLONG_FORMAT " SWPointer::offset_plus_k: Op_ConL FAILED, k is too big", n->get_long());
4179 }
4180 }
4181
offset_plus_k_5(Node * n,Node * _invar)4182 void SWPointer::Tracer::offset_plus_k_5(Node* n, Node* _invar) {
4183 if(_slp->is_trace_alignment()) {
4184 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED since another invariant has been detected before", n->_idx);
4185 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: _invar != NULL: ", _invar->_idx); _invar->dump();
4186 }
4187 }
4188
offset_plus_k_6(Node * n,Node * _invar,bool _negate_invar,int _offset)4189 void SWPointer::Tracer::offset_plus_k_6(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4190 if(_slp->is_trace_alignment()) {
4191 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4192 n->_idx, _negate_invar, _invar->_idx, _offset);
4193 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4194 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4195 }
4196 }
4197
offset_plus_k_7(Node * n,Node * _invar,bool _negate_invar,int _offset)4198 void SWPointer::Tracer::offset_plus_k_7(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4199 if(_slp->is_trace_alignment()) {
4200 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_AddI PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4201 n->_idx, _negate_invar, _invar->_idx, _offset);
4202 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4203 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4204 }
4205 }
4206
offset_plus_k_8(Node * n,Node * _invar,bool _negate_invar,int _offset)4207 void SWPointer::Tracer::offset_plus_k_8(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4208 if(_slp->is_trace_alignment()) {
4209 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: Op_SubI is PASSED, setting _negate_invar = %d, _invar = %d, _offset = %d",
4210 n->_idx, _negate_invar, _invar->_idx, _offset);
4211 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is Con: ", n->in(2)->_idx); n->in(2)->dump();
4212 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is invariant: ", _invar->_idx); _invar->dump();
4213 }
4214 }
4215
offset_plus_k_9(Node * n,Node * _invar,bool _negate_invar,int _offset)4216 void SWPointer::Tracer::offset_plus_k_9(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4217 if(_slp->is_trace_alignment()) {
4218 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);
4219 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(1) is Con: ", n->in(1)->_idx); n->in(1)->dump();
4220 print_depth(); tty->print(" \\ %d SWPointer::offset_plus_k: in(2) is invariant: ", _invar->_idx); _invar->dump();
4221 }
4222 }
4223
offset_plus_k_10(Node * n,Node * _invar,bool _negate_invar,int _offset)4224 void SWPointer::Tracer::offset_plus_k_10(Node* n, Node* _invar, bool _negate_invar, int _offset) {
4225 if(_slp->is_trace_alignment()) {
4226 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);
4227 print_depth(); tty->print_cr(" \\ %d SWPointer::offset_plus_k: is invariant", n->_idx);
4228 }
4229 }
4230
offset_plus_k_11(Node * n)4231 void SWPointer::Tracer::offset_plus_k_11(Node* n) {
4232 if(_slp->is_trace_alignment()) {
4233 print_depth(); tty->print_cr(" %d SWPointer::offset_plus_k: FAILED", n->_idx);
4234 }
4235 }
4236
4237 #endif
4238 // ========================= OrderedPair =====================
4239
4240 const OrderedPair OrderedPair::initial;
4241
4242 // ========================= SWNodeInfo =====================
4243
4244 const SWNodeInfo SWNodeInfo::initial;
4245
4246
4247 // ============================ DepGraph ===========================
4248
4249 //------------------------------make_node---------------------------
4250 // Make a new dependence graph node for an ideal node.
make_node(Node * node)4251 DepMem* DepGraph::make_node(Node* node) {
4252 DepMem* m = new (_arena) DepMem(node);
4253 if (node != NULL) {
4254 assert(_map.at_grow(node->_idx) == NULL, "one init only");
4255 _map.at_put_grow(node->_idx, m);
4256 }
4257 return m;
4258 }
4259
4260 //------------------------------make_edge---------------------------
4261 // Make a new dependence graph edge from dpred -> dsucc
make_edge(DepMem * dpred,DepMem * dsucc)4262 DepEdge* DepGraph::make_edge(DepMem* dpred, DepMem* dsucc) {
4263 DepEdge* e = new (_arena) DepEdge(dpred, dsucc, dsucc->in_head(), dpred->out_head());
4264 dpred->set_out_head(e);
4265 dsucc->set_in_head(e);
4266 return e;
4267 }
4268
4269 // ========================== DepMem ========================
4270
4271 //------------------------------in_cnt---------------------------
in_cnt()4272 int DepMem::in_cnt() {
4273 int ct = 0;
4274 for (DepEdge* e = _in_head; e != NULL; e = e->next_in()) ct++;
4275 return ct;
4276 }
4277
4278 //------------------------------out_cnt---------------------------
out_cnt()4279 int DepMem::out_cnt() {
4280 int ct = 0;
4281 for (DepEdge* e = _out_head; e != NULL; e = e->next_out()) ct++;
4282 return ct;
4283 }
4284
4285 //------------------------------print-----------------------------
print()4286 void DepMem::print() {
4287 #ifndef PRODUCT
4288 tty->print(" DepNode %d (", _node->_idx);
4289 for (DepEdge* p = _in_head; p != NULL; p = p->next_in()) {
4290 Node* pred = p->pred()->node();
4291 tty->print(" %d", pred != NULL ? pred->_idx : 0);
4292 }
4293 tty->print(") [");
4294 for (DepEdge* s = _out_head; s != NULL; s = s->next_out()) {
4295 Node* succ = s->succ()->node();
4296 tty->print(" %d", succ != NULL ? succ->_idx : 0);
4297 }
4298 tty->print_cr(" ]");
4299 #endif
4300 }
4301
4302 // =========================== DepEdge =========================
4303
4304 //------------------------------DepPreds---------------------------
print()4305 void DepEdge::print() {
4306 #ifndef PRODUCT
4307 tty->print_cr("DepEdge: %d [ %d ]", _pred->node()->_idx, _succ->node()->_idx);
4308 #endif
4309 }
4310
4311 // =========================== DepPreds =========================
4312 // Iterator over predecessor edges in the dependence graph.
4313
4314 //------------------------------DepPreds---------------------------
DepPreds(Node * n,DepGraph & dg)4315 DepPreds::DepPreds(Node* n, DepGraph& dg) {
4316 _n = n;
4317 _done = false;
4318 if (_n->is_Store() || _n->is_Load()) {
4319 _next_idx = MemNode::Address;
4320 _end_idx = n->req();
4321 _dep_next = dg.dep(_n)->in_head();
4322 } else if (_n->is_Mem()) {
4323 _next_idx = 0;
4324 _end_idx = 0;
4325 _dep_next = dg.dep(_n)->in_head();
4326 } else {
4327 _next_idx = 1;
4328 _end_idx = _n->req();
4329 _dep_next = NULL;
4330 }
4331 next();
4332 }
4333
4334 //------------------------------next---------------------------
next()4335 void DepPreds::next() {
4336 if (_dep_next != NULL) {
4337 _current = _dep_next->pred()->node();
4338 _dep_next = _dep_next->next_in();
4339 } else if (_next_idx < _end_idx) {
4340 _current = _n->in(_next_idx++);
4341 } else {
4342 _done = true;
4343 }
4344 }
4345
4346 // =========================== DepSuccs =========================
4347 // Iterator over successor edges in the dependence graph.
4348
4349 //------------------------------DepSuccs---------------------------
DepSuccs(Node * n,DepGraph & dg)4350 DepSuccs::DepSuccs(Node* n, DepGraph& dg) {
4351 _n = n;
4352 _done = false;
4353 if (_n->is_Load()) {
4354 _next_idx = 0;
4355 _end_idx = _n->outcnt();
4356 _dep_next = dg.dep(_n)->out_head();
4357 } else if (_n->is_Mem() || (_n->is_Phi() && _n->bottom_type() == Type::MEMORY)) {
4358 _next_idx = 0;
4359 _end_idx = 0;
4360 _dep_next = dg.dep(_n)->out_head();
4361 } else {
4362 _next_idx = 0;
4363 _end_idx = _n->outcnt();
4364 _dep_next = NULL;
4365 }
4366 next();
4367 }
4368
4369 //-------------------------------next---------------------------
next()4370 void DepSuccs::next() {
4371 if (_dep_next != NULL) {
4372 _current = _dep_next->succ()->node();
4373 _dep_next = _dep_next->next_out();
4374 } else if (_next_idx < _end_idx) {
4375 _current = _n->raw_out(_next_idx++);
4376 } else {
4377 _done = true;
4378 }
4379 }
4380
4381 //
4382 // --------------------------------- vectorization/simd -----------------------------------
4383 //
same_origin_idx(Node * a,Node * b) const4384 bool SuperWord::same_origin_idx(Node* a, Node* b) const {
4385 return a != NULL && b != NULL && _clone_map.same_idx(a->_idx, b->_idx);
4386 }
same_generation(Node * a,Node * b) const4387 bool SuperWord::same_generation(Node* a, Node* b) const {
4388 return a != NULL && b != NULL && _clone_map.same_gen(a->_idx, b->_idx);
4389 }
4390
find_phi_for_mem_dep(LoadNode * ld)4391 Node* SuperWord::find_phi_for_mem_dep(LoadNode* ld) {
4392 assert(in_bb(ld), "must be in block");
4393 if (_clone_map.gen(ld->_idx) == _ii_first) {
4394 #ifndef PRODUCT
4395 if (_vector_loop_debug) {
4396 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(ld->_idx)=%d",
4397 _clone_map.gen(ld->_idx));
4398 }
4399 #endif
4400 return NULL; //we think that any ld in the first gen being vectorizable
4401 }
4402
4403 Node* mem = ld->in(MemNode::Memory);
4404 if (mem->outcnt() <= 1) {
4405 // we don't want to remove the only edge from mem node to load
4406 #ifndef PRODUCT
4407 if (_vector_loop_debug) {
4408 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",
4409 mem->_idx, ld->_idx);
4410 ld->dump();
4411 mem->dump();
4412 }
4413 #endif
4414 return NULL;
4415 }
4416 if (!in_bb(mem) || same_generation(mem, ld)) {
4417 #ifndef PRODUCT
4418 if (_vector_loop_debug) {
4419 tty->print_cr("SuperWord::find_phi_for_mem_dep _clone_map.gen(mem->_idx)=%d",
4420 _clone_map.gen(mem->_idx));
4421 }
4422 #endif
4423 return NULL; // does not depend on loop volatile node or depends on the same generation
4424 }
4425
4426 //otherwise first node should depend on mem-phi
4427 Node* first = first_node(ld);
4428 assert(first->is_Load(), "must be Load");
4429 Node* phi = first->as_Load()->in(MemNode::Memory);
4430 if (!phi->is_Phi() || phi->bottom_type() != Type::MEMORY) {
4431 #ifndef PRODUCT
4432 if (_vector_loop_debug) {
4433 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");
4434 ld->dump();
4435 first->dump();
4436 }
4437 #endif
4438 return NULL;
4439 }
4440
4441 Node* tail = 0;
4442 for (int m = 0; m < _mem_slice_head.length(); m++) {
4443 if (_mem_slice_head.at(m) == phi) {
4444 tail = _mem_slice_tail.at(m);
4445 }
4446 }
4447 if (tail == 0) { //test that found phi is in the list _mem_slice_head
4448 #ifndef PRODUCT
4449 if (_vector_loop_debug) {
4450 tty->print_cr("SuperWord::find_phi_for_mem_dep load %d is not vectorizable node, its phi %d is not _mem_slice_head",
4451 ld->_idx, phi->_idx);
4452 ld->dump();
4453 phi->dump();
4454 }
4455 #endif
4456 return NULL;
4457 }
4458
4459 // now all conditions are met
4460 return phi;
4461 }
4462
first_node(Node * nd)4463 Node* SuperWord::first_node(Node* nd) {
4464 for (int ii = 0; ii < _iteration_first.length(); ii++) {
4465 Node* nnn = _iteration_first.at(ii);
4466 if (same_origin_idx(nnn, nd)) {
4467 #ifndef PRODUCT
4468 if (_vector_loop_debug) {
4469 tty->print_cr("SuperWord::first_node: %d is the first iteration node for %d (_clone_map.idx(nnn->_idx) = %d)",
4470 nnn->_idx, nd->_idx, _clone_map.idx(nnn->_idx));
4471 }
4472 #endif
4473 return nnn;
4474 }
4475 }
4476
4477 #ifndef PRODUCT
4478 if (_vector_loop_debug) {
4479 tty->print_cr("SuperWord::first_node: did not find first iteration node for %d (_clone_map.idx(nd->_idx)=%d)",
4480 nd->_idx, _clone_map.idx(nd->_idx));
4481 }
4482 #endif
4483 return 0;
4484 }
4485
last_node(Node * nd)4486 Node* SuperWord::last_node(Node* nd) {
4487 for (int ii = 0; ii < _iteration_last.length(); ii++) {
4488 Node* nnn = _iteration_last.at(ii);
4489 if (same_origin_idx(nnn, nd)) {
4490 #ifndef PRODUCT
4491 if (_vector_loop_debug) {
4492 tty->print_cr("SuperWord::last_node _clone_map.idx(nnn->_idx)=%d, _clone_map.idx(nd->_idx)=%d",
4493 _clone_map.idx(nnn->_idx), _clone_map.idx(nd->_idx));
4494 }
4495 #endif
4496 return nnn;
4497 }
4498 }
4499 return 0;
4500 }
4501
mark_generations()4502 int SuperWord::mark_generations() {
4503 Node *ii_err = NULL, *tail_err = NULL;
4504 for (int i = 0; i < _mem_slice_head.length(); i++) {
4505 Node* phi = _mem_slice_head.at(i);
4506 assert(phi->is_Phi(), "must be phi");
4507
4508 Node* tail = _mem_slice_tail.at(i);
4509 if (_ii_last == -1) {
4510 tail_err = tail;
4511 _ii_last = _clone_map.gen(tail->_idx);
4512 }
4513 else if (_ii_last != _clone_map.gen(tail->_idx)) {
4514 #ifndef PRODUCT
4515 if (TraceSuperWord && Verbose) {
4516 tty->print_cr("SuperWord::mark_generations _ii_last error - found different generations in two tail nodes ");
4517 tail->dump();
4518 tail_err->dump();
4519 }
4520 #endif
4521 return -1;
4522 }
4523
4524 // find first iteration in the loop
4525 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
4526 Node* ii = phi->fast_out(i);
4527 if (in_bb(ii) && ii->is_Store()) { // we speculate that normally Stores of one and one only generation have deps from mem phi
4528 if (_ii_first == -1) {
4529 ii_err = ii;
4530 _ii_first = _clone_map.gen(ii->_idx);
4531 } else if (_ii_first != _clone_map.gen(ii->_idx)) {
4532 #ifndef PRODUCT
4533 if (TraceSuperWord && Verbose) {
4534 tty->print_cr("SuperWord::mark_generations: _ii_first was found before and not equal to one in this node (%d)", _ii_first);
4535 ii->dump();
4536 if (ii_err!= 0) {
4537 ii_err->dump();
4538 }
4539 }
4540 #endif
4541 return -1; // this phi has Stores from different generations of unroll and cannot be simd/vectorized
4542 }
4543 }
4544 }//for (DUIterator_Fast imax,
4545 }//for (int i...
4546
4547 if (_ii_first == -1 || _ii_last == -1) {
4548 if (TraceSuperWord && Verbose) {
4549 tty->print_cr("SuperWord::mark_generations unknown error, something vent wrong");
4550 }
4551 return -1; // something vent wrong
4552 }
4553 // collect nodes in the first and last generations
4554 assert(_iteration_first.length() == 0, "_iteration_first must be empty");
4555 assert(_iteration_last.length() == 0, "_iteration_last must be empty");
4556 for (int j = 0; j < _block.length(); j++) {
4557 Node* n = _block.at(j);
4558 node_idx_t gen = _clone_map.gen(n->_idx);
4559 if ((signed)gen == _ii_first) {
4560 _iteration_first.push(n);
4561 } else if ((signed)gen == _ii_last) {
4562 _iteration_last.push(n);
4563 }
4564 }
4565
4566 // building order of iterations
4567 if (_ii_order.length() == 0 && ii_err != 0) {
4568 assert(in_bb(ii_err) && ii_err->is_Store(), "should be Store in bb");
4569 Node* nd = ii_err;
4570 while(_clone_map.gen(nd->_idx) != _ii_last) {
4571 _ii_order.push(_clone_map.gen(nd->_idx));
4572 bool found = false;
4573 for (DUIterator_Fast imax, i = nd->fast_outs(imax); i < imax; i++) {
4574 Node* use = nd->fast_out(i);
4575 if (same_origin_idx(use, nd) && use->as_Store()->in(MemNode::Memory) == nd) {
4576 found = true;
4577 nd = use;
4578 break;
4579 }
4580 }//for
4581
4582 if (found == false) {
4583 if (TraceSuperWord && Verbose) {
4584 tty->print_cr("SuperWord::mark_generations: Cannot build order of iterations - no dependent Store for %d", nd->_idx);
4585 }
4586 _ii_order.clear();
4587 return -1;
4588 }
4589 } //while
4590 _ii_order.push(_clone_map.gen(nd->_idx));
4591 }
4592
4593 #ifndef PRODUCT
4594 if (_vector_loop_debug) {
4595 tty->print_cr("SuperWord::mark_generations");
4596 tty->print_cr("First generation (%d) nodes:", _ii_first);
4597 for (int ii = 0; ii < _iteration_first.length(); ii++) _iteration_first.at(ii)->dump();
4598 tty->print_cr("Last generation (%d) nodes:", _ii_last);
4599 for (int ii = 0; ii < _iteration_last.length(); ii++) _iteration_last.at(ii)->dump();
4600 tty->print_cr(" ");
4601
4602 tty->print("SuperWord::List of generations: ");
4603 for (int jj = 0; jj < _ii_order.length(); ++jj) {
4604 tty->print("%d:%d ", jj, _ii_order.at(jj));
4605 }
4606 tty->print_cr(" ");
4607 }
4608 #endif
4609
4610 return _ii_first;
4611 }
4612
fix_commutative_inputs(Node * gold,Node * fix)4613 bool SuperWord::fix_commutative_inputs(Node* gold, Node* fix) {
4614 assert(gold->is_Add() && fix->is_Add() || gold->is_Mul() && fix->is_Mul(), "should be only Add or Mul nodes");
4615 assert(same_origin_idx(gold, fix), "should be clones of the same node");
4616 Node* gin1 = gold->in(1);
4617 Node* gin2 = gold->in(2);
4618 Node* fin1 = fix->in(1);
4619 Node* fin2 = fix->in(2);
4620 bool swapped = false;
4621
4622 if (in_bb(gin1) && in_bb(gin2) && in_bb(fin1) && in_bb(fin1)) {
4623 if (same_origin_idx(gin1, fin1) &&
4624 same_origin_idx(gin2, fin2)) {
4625 return true; // nothing to fix
4626 }
4627 if (same_origin_idx(gin1, fin2) &&
4628 same_origin_idx(gin2, fin1)) {
4629 fix->swap_edges(1, 2);
4630 swapped = true;
4631 }
4632 }
4633 // at least one input comes from outside of bb
4634 if (gin1->_idx == fin1->_idx) {
4635 return true; // nothing to fix
4636 }
4637 if (!swapped && (gin1->_idx == fin2->_idx || gin2->_idx == fin1->_idx)) { //swapping is expensive, check condition first
4638 fix->swap_edges(1, 2);
4639 swapped = true;
4640 }
4641
4642 if (swapped) {
4643 #ifndef PRODUCT
4644 if (_vector_loop_debug) {
4645 tty->print_cr("SuperWord::fix_commutative_inputs: fixed node %d", fix->_idx);
4646 }
4647 #endif
4648 return true;
4649 }
4650
4651 if (TraceSuperWord && Verbose) {
4652 tty->print_cr("SuperWord::fix_commutative_inputs: cannot fix node %d", fix->_idx);
4653 }
4654
4655 return false;
4656 }
4657
pack_parallel()4658 bool SuperWord::pack_parallel() {
4659 #ifndef PRODUCT
4660 if (_vector_loop_debug) {
4661 tty->print_cr("SuperWord::pack_parallel: START");
4662 }
4663 #endif
4664
4665 _packset.clear();
4666
4667 for (int ii = 0; ii < _iteration_first.length(); ii++) {
4668 Node* nd = _iteration_first.at(ii);
4669 if (in_bb(nd) && (nd->is_Load() || nd->is_Store() || nd->is_Add() || nd->is_Mul())) {
4670 Node_List* pk = new Node_List();
4671 pk->push(nd);
4672 for (int gen = 1; gen < _ii_order.length(); ++gen) {
4673 for (int kk = 0; kk < _block.length(); kk++) {
4674 Node* clone = _block.at(kk);
4675 if (same_origin_idx(clone, nd) &&
4676 _clone_map.gen(clone->_idx) == _ii_order.at(gen)) {
4677 if (nd->is_Add() || nd->is_Mul()) {
4678 fix_commutative_inputs(nd, clone);
4679 }
4680 pk->push(clone);
4681 if (pk->size() == 4) {
4682 _packset.append(pk);
4683 #ifndef PRODUCT
4684 if (_vector_loop_debug) {
4685 tty->print_cr("SuperWord::pack_parallel: added pack ");
4686 pk->dump();
4687 }
4688 #endif
4689 if (_clone_map.gen(clone->_idx) != _ii_last) {
4690 pk = new Node_List();
4691 }
4692 }
4693 break;
4694 }
4695 }
4696 }//for
4697 }//if
4698 }//for
4699
4700 #ifndef PRODUCT
4701 if (_vector_loop_debug) {
4702 tty->print_cr("SuperWord::pack_parallel: END");
4703 }
4704 #endif
4705
4706 return true;
4707 }
4708
hoist_loads_in_graph()4709 bool SuperWord::hoist_loads_in_graph() {
4710 GrowableArray<Node*> loads;
4711
4712 #ifndef PRODUCT
4713 if (_vector_loop_debug) {
4714 tty->print_cr("SuperWord::hoist_loads_in_graph: total number _mem_slice_head.length() = %d", _mem_slice_head.length());
4715 }
4716 #endif
4717
4718 for (int i = 0; i < _mem_slice_head.length(); i++) {
4719 Node* n = _mem_slice_head.at(i);
4720 if ( !in_bb(n) || !n->is_Phi() || n->bottom_type() != Type::MEMORY) {
4721 if (TraceSuperWord && Verbose) {
4722 tty->print_cr("SuperWord::hoist_loads_in_graph: skipping unexpected node n=%d", n->_idx);
4723 }
4724 continue;
4725 }
4726
4727 #ifndef PRODUCT
4728 if (_vector_loop_debug) {
4729 tty->print_cr("SuperWord::hoist_loads_in_graph: processing phi %d = _mem_slice_head.at(%d);", n->_idx, i);
4730 }
4731 #endif
4732
4733 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4734 Node* ld = n->fast_out(i);
4735 if (ld->is_Load() && ld->as_Load()->in(MemNode::Memory) == n && in_bb(ld)) {
4736 for (int i = 0; i < _block.length(); i++) {
4737 Node* ld2 = _block.at(i);
4738 if (ld2->is_Load() && same_origin_idx(ld, ld2) &&
4739 !same_generation(ld, ld2)) { // <= do not collect the first generation ld
4740 #ifndef PRODUCT
4741 if (_vector_loop_debug) {
4742 tty->print_cr("SuperWord::hoist_loads_in_graph: will try to hoist load ld2->_idx=%d, cloned from %d (ld->_idx=%d)",
4743 ld2->_idx, _clone_map.idx(ld->_idx), ld->_idx);
4744 }
4745 #endif
4746 // could not do on-the-fly, since iterator is immutable
4747 loads.push(ld2);
4748 }
4749 }// for
4750 }//if
4751 }//for (DUIterator_Fast imax,
4752 }//for (int i = 0; i
4753
4754 for (int i = 0; i < loads.length(); i++) {
4755 LoadNode* ld = loads.at(i)->as_Load();
4756 Node* phi = find_phi_for_mem_dep(ld);
4757 if (phi != NULL) {
4758 #ifndef PRODUCT
4759 if (_vector_loop_debug) {
4760 tty->print_cr("SuperWord::hoist_loads_in_graph replacing MemNode::Memory(%d) edge in %d with one from %d",
4761 MemNode::Memory, ld->_idx, phi->_idx);
4762 }
4763 #endif
4764 _igvn.replace_input_of(ld, MemNode::Memory, phi);
4765 }
4766 }//for
4767
4768 restart(); // invalidate all basic structures, since we rebuilt the graph
4769
4770 if (TraceSuperWord && Verbose) {
4771 tty->print_cr("\nSuperWord::hoist_loads_in_graph() the graph was rebuilt, all structures invalidated and need rebuild");
4772 }
4773
4774 return true;
4775 }
4776