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